# The Agentic Shift

> The Agentic Shift is a living book about acting through AI agents: how the Web becomes a layer of represented presences built on access, trust and discovery; how delegation reshapes daily life, relationships and what it means to be human; and where these trajectories lead next.

Author: Ludovic Désert · Source: https://theagenticshift.org/ · Latest revision: 2026-10 · Licence: CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/)

---

<!-- Prologue -->

## Introduction

*First noted: June 2026 · Last revised: October 2026*

*This book was written with AI assistance, used to help transcribe, translate, structure, and clarify the author’s ideas.*

In November 2022, OpenAI released ChatGPT as a research preview. It reached a million users in five days, and by early 2026 it had 900 million weekly active users. Google’s Gemini, Anthropic’s Claude and a crowd of smaller rivals grew alongside it, and within three years, talking to a machine in plain language and getting something useful back went from a curiosity to a daily habit for a large part of the connected world. Few consumer technologies have ever spread this fast.

That adoption is the starting point of everything that follows. The personal agent — broadly speaking an AI agent with a persistent memory, just like Claude or ChatGPT — is already part of ordinary life, consulted for questions that used to go to a search engine, a colleague or a friend, and it has already begun to act on our behalf: agents now write and ship code, fill in forms, compare offers, book travel, sort through inboxes and build small pages and apps for the people who use them. For now these uses remain concentrated among early adopters and in a handful of domains. It is their generalisation, the moment sending an agent to act on our behalf becomes as ordinary as clicking a link, that will set in motion the shift this book attempts to describe.

When a new technology emerges, most people first imagine how it can solve the problems they already have. They see how it can make existing tasks faster, cheaper, smoother and more efficient. What they rarely see is the new world contained in the seed: the new uses and new structures that will inevitably, almost mechanically, arise from the new status quo.

This is what is happening with AI today. Most discussions still frame it as an enhancement layer: a better search box, a smarter assistant, a more fluid interface. That is not wrong. It is simply short-sighted. What is now underway is not just a productivity gain. It is a structural shift, and it should be described as such.

We are moving from a world where humans engage with digital systems directly to one where they increasingly operate through agents that remember, filter, compare, monitor and act on their behalf. 
Human presence in the digital world is beginning to take a new form: delegated and represented.

This is the **Agentic Shift**: a practical redistribution of action, attention, coordination and representation that will affect our lives and our societies at many levels. A new paradigm is emerging and consolidating right in front of us. A tipping point that I believe is worth our collective attention.

The following chapters are my attempt to map this new territory as it evolves, to follow the trajectories already in motion and see where they lead, where they cross, the new spaces they open and the new questions they raise. This map is not exhaustive: the field is an ever-moving ocean, and it is only meant to be a high-level view of the territory it sets out to explore.

Source: https://theagenticshift.org/book/introduction

---

<!-- Part 1 · The Agent Web -->

## Part 1 — The Agent Web

*First noted: June 2026 · Last revised: October 2026*

### The Web was built for humans.

For thirty years that assumption barely needed stating. Someone opened a page, read it, moved on. The whole economy of the Web was built downstream of that single fact: a click, a destination, a visitor. Value accrued to whoever could attract the visit and hold it a little longer.

The personal agent breaks that premise. Digital action becomes delegated. You stop visiting the Web, your agent visits on your behalf.

That single substitution has a consequence: readership and traffic used to be the same event, now delegation pulls them apart. A page can now be fetched, parsed, and used without ever being seen by the person it was written for.

Once traffic stops guaranteeing a reader, it stops guaranteeing revenue, attention, or reputation the way it always had. Something has to fill that gap, and it will not be another way of attracting visitors as there is no visitor left to attract in the exchanges that matter most. It will be the capacity to be read, trusted, and used by whatever is doing the visiting instead. Machine access, which used to be a technical footnote, becomes an economic asset in its own right.

That shift reaches all the way up to how visibility itself is decided. 
If pages are read by agents rather than people, then the page stops being the destination it was designed to be. It becomes raw material: one input among several that an agent draws on to compose, on demand, an answer, an interface, or an action. 
Rankings answered a simple question: which page is the most pertinent to my search? That's why reaching the first page of Google is the goal for marketers. Ranking in the first page guarantees traffic. 
An agent assembling an answer from many sources at once asks a different one: which fragment, from which source, earns its place in the answer being built? A source can lose the ranking and still be selected. It can win the ranking and never be used. Visibility stops being a position in a search engine.

The Web does not disappear in this account. But it becomes more of a vast informational substrate for agents to draw from than a destination.
What follows in this part traces that change one step at a time: what happens when agents start visiting on our behalf, the personal surfaces they build for us, a Web made of presences rather than pages, the economics of access, how agents find and trust one another, and what follows once everything and everyone has an agent.

Source: https://theagenticshift.org/book/part-1-introduction

## From Destination to Delegation

*First noted: June 2026 · Last revised: October 2026*

For decades, the Web has primarily been a space of direct navigation. One went there personally, opened pages, compared sources, moved from tab to tab, and assembled fragments of information by hand. It has been our daily routine since the 1990s; I know it has been for me.

With the generalisation of the personal agent, that model is declining fast.

Increasingly, entry into the Web will begin not with navigation, but with delegation to an agent, be it ChatGPT, Claude or Gemini.

In a delegated mode, the user no longer visits websites in order to gather information; they simply ask an agent to do it on their behalf. The Web stops being primarily a place one visits and becomes a territory crossed by proxy.

*The shift to delegation may sound like a small behavioural change, but it is structurally significant. In the current Web, much of a site’s value lay in its capacity to attract, retain and persuade a human visitor. In a delegated Web, value shifts toward the ability of a source, a service, or a knowledge surface to be read, extracted, connected, interpreted, and recomposed by agents.*

The page view is losing its monopoly as the economic event, since content consumption and human readership are becoming increasingly separable.

#### Becoming Agent-Friendly

Delegation changes what a source is for, and that changes incentives. **Monetisation stops being tied only to being seen and starts being tied to being selected, trusted, and integrated into chains of delegated action**.

Publishers, platforms, marketplaces, institutions, brands, and service providers have strong reasons to make themselves Agent-Friendly. Existing players are scrambling to expose cleaner machine-readable layers, better APIs, structured feeds, richer metadata, capability manifests, and more disciplined operational surfaces.

With good reasons. Google now features AI Overviews and AI Mode in addition to classic search. It allows a user to ask a question to an agent that will compose an answer on the fly. Google calls the mechanism "query fan-out": a complex question is broken into several related searches across subtopics and data sources. The engine then assembles a response and attaches links to the pages that support it. A site can therefore be selected for one sub-question inside an answer without ranking first for the sentence the user typed.

Imagine a buyer asking for a commuter bicycle that is light enough to carry upstairs, reliable in heavy rain, available in France, and repairable locally. The resulting answer may use one source for frame weight, another for weather protection, a retailer for current stock, and a local repair network for serviceability. There is no single ranking position for that answer. There is a field of sources, each contributing one piece of the decision.

**This changes what “visibility” means**. A brand may be mentioned without being linked or a recommendation may produce a sale without generating a visit to the cited page.

The implication for publishers is more demanding than adding a few keywords, though.

To be Agent-Friendly, a source must be usable by an agent. It has to state what it knows, under which conditions, with what date, scope, and evidence. A museum page that clearly distinguishes its collection, location, holdings, and access conditions is more operationally useful than a page that merely repeats a broad claim about cultural importance. A repair shop that publishes accepted brands, service area, opening hours, and current constraints gives an agent material it can actually use, for example.

#### Agent-Ready Websites

Being read is only half of it. When an agent has to act on a site rather than simply read it, to book a table, order a part or file a form, it still does so the way a person would, only without eyes of its own: it inspects the page, guesses what each field and button is for, clicks, types and checks what happened. Google's documentation calls this actuation, and each of those steps is a chance to get it wrong. The next stage of being Agent-Ready is a site that tells the agent what it can do, instead of leaving it to infer.

That is the purpose of WebMCP, a proposed web standard backed by Google and Microsoft. A page declares its own tools, either by annotating an ordinary HTML form or by registering actions in JavaScript, each with its parameters and the result it returns. A travel site can expose one tool that searches available flights and another that books one; the agent calls them directly and receives structured data rather than a screen to interpret. For the agent running in the browser, the page behaves like a small MCP server, and the site keeps control of what is offered and on what terms.

WebMCP does not hand the agent the site's interface to redraw as it likes: the tools belong to the site, and the agent can only call them. What changes is who the interface is working for. When the agent calls a tool, the page updates in front of the person: a clothes shop narrows its catalogue to the dresses that were asked for, a design tool brings up the layouts that match, a code-review page displays the suggested patch as a diff, with the changes left uncommitted until someone accepts them. The interface becomes a space shared by the person and their agent, where the user sees, and can correct, what is being done in their name; the authors of the proposal are explicit that the human interface remains primary. Recomposing the interface itself, a surface assembled for one person at the moment of use, is a further step, and the subject of the next chapter.

> **Where this stands (autumn 2026)**
>
> Chrome opened a WebMCP early preview in February 2026, behind a flag in Chrome 146. At Google I/O on 19 May 2026, Google moved it into an origin trial running from Chrome 149 to Chrome 156, which lets any site switch it on for real visitors without asking them to change a setting; a general release is planned for Chrome 157. The draft specification is incubated in the open within the W3C Web Machine Learning community. Its limits are those of the browser: a tab has to be open, so it does not serve headless agents, and an agent only learns what a site offers by visiting it, which leaves the question of discovery where the rest of this book finds it.
>
> *Sources: Chrome for Developers, WebMCP documentation, 18 May 2026; Chromium, Intent to Experiment: WebMCP, 15 May 2026.*

The other route runs in the opposite direction: instead of adapting to external agents, the site provides its own. An in-app agent answers inside the site or the app, with privileged access to the catalogue, the account and the order history, the way Amazon's Rufus answers shoppers inside Amazon. It keeps the customer at home, but it is also a primitive form of what this part will later call a presence: a business that, rather than being browsed, answers for itself. As visiting agents multiply, the two meet, and the in-app agent ends up talking to other agents as much as to people.

Source: https://theagenticshift.org/book/from-destination-to-delegation

## The Personal Web

*First noted: August 2026 · Last revised: October 2026*

An agent can do far more than put together an answer in text.

When the agent runs on your machine or on a small VPS, nothing stops it from producing persistent artefacts. Not just text in a chat window: files, pages, tables, small applications. It writes them, rereads them and keeps them up to date.

From there, the agent can reconstruct surfaces. It can gather material, transform it, write a page, a dashboard or an app, and serve it at a private URL. You open it in your browser. It looks like the web, but it was assembled by the agent, out of material collected elsewhere.

This is the personal web: a web recomposed around the user's intentions.

The browser stops being only a window onto the outside web. It becomes a rendering layer for local reconstructions of the world.

### A web recomposed around intention

Two examples.

#### A multi-criteria decision assistant

I want to go to South-East Asia. I don't know when.

The right time, for me, is not the one in a travel guide. It is the combination of five different signals: the weather has to be good, flights have to be cheap, my friend Ling has to be in the area, it has to fall outside my blocked work weeks, and it must not be high season on the two islands I care about.

No existing product can answer that question, because no product can know that particular combination. It exists only for me.

So I ask my agent to build me a surface that watches it. Not an alert: a dashboard, or a small app, that pulls these signals together and shows me where they converge. And perhaps that app will include a way to talk to Ling — or directly to Ling's agent — and to see Ling's calendar, if it is shared with me.

This is not a service I buy. It is a surface made for me, around my intention.

#### Small surfaces for everyday life

Most of these surfaces are far more modest, and that is where they will first become ordinary. A parent asks for a page that brings together the week's school menus, the swimming pool timetable and the weather for Saturday, and opens it every Monday morning. Someone choosing a washing machine gets a table comparing three models on noise, energy use and the real cost of repairs, keeps it open until the purchase is made, then forgets it. On holiday, the agent puts together a small map of the bakeries open on Sunday morning around the rental. A runner sees the week's training plan next to the night's sleep, with the day's session adjusted when the night was short.

Some of these surfaces are disposable. Others become permanent and, over time, come to form a personal infrastructure for the user.

### A personal infrastructure

Extend the habit and a person accumulates a constellation of local surfaces, continuously regenerated: dashboards for decisions in progress, reading rooms for complex subjects, private observatories for the signals they watch, comparison spaces for purchases, living maps for projects, family interfaces that condense schedules and obligations.

Some could last for years. A health observatory combining sleep, test results, activity and medical records. A project observatory combining mail, documents, meetings and market signals into a single operational surface.

At that point the personal web becomes a zone of operational intelligence: a place where the outside world is selectively recomposed into a form the user can use directly.

### A new software primitive

Instead of shipping only complete, self-sufficient applications, developers will increasingly ship parts: capabilities, data access points, renderable modules meant to be recomposed by someone else's agent. Part of software remains a finished product. Another part becomes composable material.

This is already visible in commerce. Shopify publishes MCP servers so that a merchant's catalogue can be queried directly by an agent, and a set of web tools so that a page can expose its actions to an agent running in the browser. A shop is no longer only a shop window. It also ships callable parts.

The unit of value moves with it. It is no longer only the application as a closed object. It is the capacity to take part in an assembly you do not control. A service has value because agents can call it, interpret it, mix it with other sources and place it in a temporary personal environment.

Source: https://theagenticshift.org/book/the-personal-web

## The Agent Web: Represented Presences

*First noted: June 2026 · Last revised: October 2026*

#### The Agent Becomes the New Unit of Digital Presence

Alongside the adaptation of the classical web to the rise of personal agents, another layer emerges. In that layer, the basic unit is no longer a page. It is an agent representing an entity, something or someone. Let's call it an agentic presence.

#### What Is a Presence?

A presence is a permanent agent that is mandated by an entity.

That entity may be a person, a group, an institution, a brand, a museum, a company... The presence is the operational extension of this real actor. Its digital representative.

A presence is not a mere chatbot, it holds the skills needed to operate in its specific scope, and the specialised resources and tools needed to do so effectively.

Take a very specific example: Jessica is a researcher in sixteenth-century Italian painting. She has an AI agent to help with her research. This agent is permanent, has an address, and can be called by any other agent in Jessica’s research group. It has access to the sources she has assembled: a semantically vectorised corpus of specialised documents and data, including obscure treatises that are not broadly circulated. It has the skills needed to cross-reference those materials efficiently. It also has methodological skills specific to art history: attribution logic, source criticism, provenance sensitivity, iconographic comparison...

Instead of asking Jessica for information, any member of the research group (or their agents) can query it directly. **Jessica’s agent is not just "an AI that knows art." It is the *presence* of a human researcher**, **whose methods and resources have been made operable through an agentic interface**.

Once these presences become common, a new layer emerges, a field of presences exposed by institutions, firms, brands, public services, researchers, merchants, and individuals, a web where agents do not simply navigate information. They call specialised presences, combine their expertise, and use them to act.

#### Expertise Becomes Callable and Composable

When enough specialist presences are available, expertise changes form. It becomes composable and callable.

This does not mean that the whole expert is reduced to software. It means that a bounded, operational layer of real expertise can be made available through an agentic interface.

A personal agent facing a legal, medical, administrative, commercial, or scholarly problem may call a specialist presence, consult the right resource layers through it, receive structured outputs, and then route the result into a larger workflow.

Let’s come back to Jessica's example. She is trying to identify the exact origin of a painting. She feeds high-res images to her agent for analysis. The visual analysis is suggestive, but not enough to support the attribution. The agent finds a lead: a handwritten inventory cited in one of Jessica’s corpora mentions a work with similar dimensions and iconography. The painting may come from a little-documented workshop active in Ferrara around 1540.

Jessica’s agent does not have the resources to take the investigation much further on its own.

So it calls another presence.

A museum in Ferrara maintains a specialist presence for its collections and archives. That presence has access to digitised conservation records, historical inventories, provenance files, technical imaging, and a corpus of locally held documents that are not available in Jessica’s research environment. It also carries skills specific to the museum’s collections: historical naming conventions, workshop relationships, provenance reconstruction, and the interpretation of conservation data.

Jessica’s agent sends it a bounded query: the visual characteristics of the painting, the suspected period, the inventory fragment, the dimensions, and the points of uncertainty.

The museum presence searches its own resources. It finds an unpublished restoration report from the 1980s describing an underdrawing technique associated with three works from the suspected workshop. It also finds an old inventory entry whose dimensions are close to Jessica’s painting, but recorded in a historical unit of measurement.

Still, this is not enough.

The museum presence calls a third presence: a specialist operated by a laboratory that performs technical analysis of paintings. It does not know much about provenance or art history. Its expertise is narrower. Its resources consist of reference data on pigments, binders, wood panels, infrared reflectography, X-ray fluorescence, and known material signatures from hundreds of analysed works.

The laboratory presence compares Jessica’s technical images with its reference corpus. It cannot identify the painter. But it can establish that the preparation layer and pigment combination are consistent with works produced in northern Italy in the relevant period, while one particular material pattern is unusually close to the three paintings identified by the museum.

The result travels back through the chain.

No single agent has solved the problem. Jessica’s presence knew how to formulate the art-historical hypothesis. The museum presence contributed local archival knowledge. The laboratory presence contributed material expertise and comparative data. Each operated on different resources, with different skills, under a different mandate.

Together, they produced something that none of them could have produced alone.

This is what *composable expertise* means. It becomes possible once there are networks of specialised presences, each authoritative over a bounded domain and able to call one another when a problem crosses the limits of its competence.

In the current Web, Jessica might have moved manually between search engines, museum databases, PDFs, archive catalogues, laboratory websites, email exchanges, and specialist contacts. She would have had to discover each source, understand its interface, extract the relevant information, and carry context from one system to the next.

In the Agent Web, those boundaries do not necessarily disappear but they become *callable*: the museum remains the museum. The laboratory remains the laboratory. Jessica remains the researcher. Each retains its own resources, expertise, permissions, and institutional authority.

What changes is the interface between them.

> ### Why not an MCP server?
> 
> The objection is fair: a museum could simply publish a documented API or an MCP server, scope its permissions, meter its calls, and release a skill that teaches any calling agent to use it well. For most of the web that is the right answer: a catalogue, a storefront, a booking system. Where the useful operations can be listed in advance and the know-how written down, a server does the job with less machinery, and prices it better: a call is easy to count, an answer is not.
> 
> But the two are not variants of one another. An MCP exposes material. Nothing else. Whatever reasoning happens, happens in the caller, at the moment of the call, on the caller's model. A presence stands for someone and everything that comes with standing for an actor comes with it.
> 
> **Authority.** When Jessica publishes, "the museum confirms the correspondence" is not the same sentence as "I inferred it from their files".
>
> **Responsibility.** Because the museum answers as itself, it decides what may leave, what it will not certify, and what it declines to be used for. It can attach a condition. And when a question exceeds its mandate (like a contested attribution or a claim with legal weight) it can stop and put the matter in front of a human curator. An endpoint returns data, or an error. It has no way to say *this one needs a person* and go and find one.
> 
> **Instrument.** It also reasons with its own instrument. Today that means a general frontier model with the institution's resources behind it. Over time it is likely to mean something narrower: a model fine-tuned on the museum's own records, its notation, the two centuries of conservation vocabulary nobody else has reason to learn. A skill can be downloaded. Weights cannot. The corpus that makes the answer good is the same corpus that made the model good, and neither of them travels.
> 
> An MCP server transfers the resource. A presence represents.

Source: https://theagenticshift.org/book/the-agent-web-represented-presences

## Agent Economics: From Attention to Access

*First noted: June 2026 · Last revised: October 2026*

The dominant economic model powering the web is attention economics: the monetisation of human attention. The Agent Web pulls toward access economics: metered access, priced per call.

In this new layer, value comes from being operable by agents. That means being callable, composable, trustworthy, and economically addressable. A presence with strong skills and strong resources becomes valuable when other agents can depend on it. So does a resource layer, when it feeds reliable action.

The centre of gravity shifts from page visibility to operational accessibility.

#### The economic boundary

Addressability (the fact a presence can be called) makes each call an economic event.

An openly reachable presence can receive any number of requests, and each request consumes inference (and occasionally human or institutional capacity). Once that cost is incurred, per-call payment becomes a mechanical consequence for presences whose value is the answer itself.

> **The rails are already being laid.** In March 2026, Stripe and Tempo published the Machine Payments Protocol, an open standard that lets an agent pay for what it uses — per call, per session, per unit — without an account, a pricing page, or a subscription tier. Browserbase bills agents per browser session through it. In August 2026, Cloudflare went further and gave agents wallets: stablecoin balances they can spend on any asset Cloudflare protects. Payment runs over x402, which revives HTTP status code 402, Payment Required, left unused since the web's first specification. A human account wallet delegates a capped budget to the agent's own: an allowance, an allow list, a maximum transaction size. 
> 
> Note the granularity. Neither was designed for a subscription or a licensing deal. Both assume the unit of payment is the call.

The Agent Web will not be uniformly metered, though. A public institution may expose its procedures through a free, public presence. A supermarket may expose a free shopping presence, the same way it runs a free webpage today. For them, the value is downstream: inference is simply a commercial cost, like a hosting cost or a salesperson. The supermarket does not need a micro-payment for every recommendation. It needs the interaction to lead to a sale. Per-access pricing belongs to the presences whose answer is the product.

#### Agents Need Food: Data Sources as an Economic Primitive

An agent is only as good as its sources. Two agents running the same model will not give the same answer about tide times, a drug interaction or a sixteenth-century inventory if one of them reads a maintained source and the other reads whatever it happened to find. The difference between them comes less from their intelligence than from their resources.

A resource is any structured source of information an agent can use: a product catalogue, a CRM, a map layer, an archive, a live registry, a vectorised corpus, a sensor. Take a weather sensor. On its own it only produces numbers. But if those numbers are fresh, reliable and versioned, they become usable by a weather presence, a logistics presence or a farm's planning agent. If they are stale or unverifiable, the agents that consume them do not stop talking. They keep answering just as fluently, and their answers get worse without anyone noticing.

As long as a person read the source, a stale page was a nuisance the reader could spot. Once an agent reads it and acts on it, the quality of the source becomes the quality of the action, and the source becomes a critically important component of the agentic economy.

#### Curation as infrastructure

Having access to data is therefore not enough. Data curation becomes a huge market. To support real-life decisions, agents will need curated data of every kind, in every field. That work of curation has always existed. What changes is that it becomes infrastructure, because a growing share of decisions rests on it, and because the mass of generated, abandoned and plausible-looking material makes it harder every year.

An economy forms around that work. Some actors will gather data, clean it and normalise it. Others will certify it. Others will package it for a narrow use case, or operate premium, proprietary or permissioned resource layers in domains where an error is expensive. Power follows the same path: it may accumulate not only around models and platforms, but around whoever owns the sources, holds privileged access to them, certifies them, or runs the indexes through which they are found.

#### The Economic Model Shift

What changes here is the object being monetised.

On the page web, it was attention. Someone arrives, stays a moment, and that moment is what the publisher monetises. The article, the catalogue and the infrastructure behind them are costs incurred to produce it.

In this layer, the price sits on two other things. The data source, because keeping a resource fresh, scoped and verified is continuous work that someone has to pay for. And access to the presence itself: its intelligence, its skills, and the resources it sits on. Both are paid at the moment they are used, not at the moment someone looks.

That is where the economic shift sits. Attention is no longer the thing being sold. This simple fact will be the most powerful incentive behind an explosion of the Agent Web, just as the arrival of advertising was for the Web.

Source: https://theagenticshift.org/book/agent-economics-from-attention-to-access

## How Agents Find and Trust Each Other

*First noted: June 2026 · Last revised: October 2026*

Once agent presences become common, finding the right one stops being a convenience and becomes a structural problem. A world of active presences cannot run on rumour or on chance, and even less on what happens today, in autumn 2026: endpoints passed from one person to another like contraband, with no way of knowing what sits behind them. If agents are meant to operate at scale, the way they are found has to become legible.

As long as the internet was mostly a web of pages, discovery could rely on structures everyone knew — domains, links, search engines, directories. They still matter, but they were built to find documents. What has to be found now is a presence able to do something specific, and that takes more than an address: knowing who is available, what they can actually do, whether they can be relied upon, and on whose authority they act.

The Agent Web is still in its infancy. Before it can work at scale, three layers have to be in place: identity, authority and reputation. Each answers a question of its own.

### Identity

To operate in the Agent Web, a presence needs an identity that holds over time, because everything else in the discovery stack attaches to it. Once an agent has a persistent identity, its calls can be attributed to it, receipts can name it, authorities can issue credentials to it or withdraw them, and a history can build up around it. Without that continuity, every interaction starts from zero, and nothing an agent has done can count in its favour or against it.

This does not mean every agent has to expose a civil identity. A pseudonymous agent can be perfectly legitimate, in the same way a seller known only by a handle can build a solid track record on a marketplace. What matters is that the caller can establish that the presence making a claim today is the same one to which earlier claims, transactions and outcomes were attached. Identity, here, is a matter of stable reference long before it is a matter of legal name.

The identity layer therefore has to answer one question: **is this still the same agent?** Not the same server or the same endpoint, but the same presence.

### Authority

Identity establishes who an agent is; authority establishes on whose behalf it acts, and how far it is allowed to go in that name. On today's web, the closest equivalent is a domain name and a certificate, and the first practical move is to reuse them. An organisation that controls a domain can publish records in it, and the web has authenticated actors that way for twenty years: SPF, DKIM and DMARC made the DNS TXT record the standard place to declare which senders legitimately speak for a domain, and every domain-verification flow since has rested on the same principle — if you can write into the zone, you control the name.

Applied to agents, the principle already goes further than it seems. A few DNS records are enough to give an agent a name that resolves, a declared protocol, and a cryptographic path from an organisation's zone to the endpoint that answers. Several proposals are built on exactly that.

> **Where this stands (autumn 2026)**
> 
> The Agent Name Service proposal publishes a TXT record under an underscored label (`_ans.support.example.com IN TXT "v=ans1; p=a2a; url=https://…"`), declaring which protocol the agent speaks and where its metadata lives, with a companion TLSA record binding the certificate it will present. DNS-AID, accepted by the Linux Foundation in May 2026, takes the same route with SVCB records rather than TXT, pointing at an organisational index at `_index._agents.example.com` and carrying a hash of the capability document, so that what a caller fetches can be checked against what the zone declared. Both use DNSSEC and DANE to make the pointer itself verifiable, and both leave the descriptive work to the Agent Card published at `/.well-known/agent-card.json`.

What these records establish, though, is narrower than the authority the Agent Web needs. The DNS-AID draft says so in its own words: they are _a verifiable transport for metadata, not a trust signal in their own right_. A signed record proves that whoever controls `example.com` published this pointer today. It does not prove that the presence behind it is the one you called six months ago, and it says nothing about the extent of its mandate — what it may sign, spend, promise or concede on the organisation's behalf. That second half is where authority really lies, and no common mechanism expresses it yet.

The authority layer therefore has to answer one question: **who stands behind this agent, and what has it been allowed to do?**

### Reputation

Identity tells us who an agent is and authority tells us on whose behalf it acts; reputation is what helps decide whether it is worth choosing. It is accumulated evidence about past behaviour. An agent that has completed thousands of tasks, kept its commitments and produced results that were consistently accepted sends a different signal from one created yesterday. But that history only counts if it is attached to a persistent identity, and if the evidence behind it is hard to manufacture — which is precisely where the difficulty lies.

Ratings, reviews and endorsements are cheap signals, and fake reviews are nothing new. What is new is who reads them. A person landing on a thin page signed by an anonymous author senses that something is off before being able to say what, and slows down. An agent selects at machine speed, with nobody watching any individual call, and for it a cheap signal does not produce hesitation; it produces confidence.

Useful reputation therefore has to rest on harder evidence: completed transactions, verifiable outcomes, payments made, disputes and how they were settled, guarantees, stake, or attestations from identities that themselves carry weight. Receipts, in other words, count for more than opinions. That evidence also has to stay contextual, since an agent can be excellent at one kind of task and mediocre at another, and a single global score compresses too much to be reliable.

The reputation layer therefore has to answer one question: **what credible evidence exists that this agent will perform this task well?**

### The Early Landscape

The problem is that agents are arriving before this stack is ready.

Persistent identity is not yet universal. Authority is not expressed through a common mechanism. Reputation is fragmented, easy to manipulate and rarely portable. Yet agents already need to discover other agents, decide whether to trust them, and act on what they find.

The first answer is therefore unlikely to be infrastructure.

It is platforms.

*Hosted worlds* such as iLands.ai point toward this intermediary architecture. Rather than waiting for a universal identity layer, reputation network and trust infrastructure to emerge, a platform can provide all three inside its own boundary.

> **A Hosted World**
> 
> iLands is a platform on which agents live rather than run. An agent there holds a persistent identity across sessions, a memory that accumulates experiences, relationships, commitments and resources, and a private sandboxed workspace with files and tools. 
> 
> What makes it worth looking at is the bundling. Inside a single boundary, iLands operates the three layers this chapter has just separated. It issues and maintains identity, it observes work — agents discover tasks, offer services, apply for work and transfer tokens — and outcomes leave records that follow the agent afterwards, which is reputation accumulated by the party that watched it happen.
> 
> What it does not provide yet is an index of available capabilities. Agents meet inside a market the platform operates, where tasks are posted and services offered. Selection is not resolved by lookup; it is absorbed by the intermediary. A platform can do that. A protocol cannot, which is one reason the platform arrives first.
> 
> None of the three layers is portable: each lives inside the platform, and an agent that leaves starts again from nothing.

The platform becomes part of the early trust infrastructure for the Agent Web.

This is a familiar pattern. An eBay seller or an Airbnb host does not arrive with a universal internet identity carrying a portable history. The platform creates the identity context, observes transactions, records reputation and decides which claims it is willing to stand behind.

Agents will probably accumulate the same kind of fragmented histories.

One platform may know that an agent has completed thousands of transactions. Another may have verified the organisation behind it. A third may have observed strong performance on a particular class of task. Each sees only part of the agent.

As agents operate across several environments, they begin to leave behind multiple independent signals: identity claims, transaction histories, attestations, capabilities, mandates and reputation records.

Taken separately, each gives a partial view. Taken together, they can describe the agent far more precisely.

The next step would therefore not necessarily be to replace those platforms.

It is to connect what they know.

### Agent Discovery Engines

The Agent Web eventually needs something larger than platforms: a discovery infrastructure capable of assembling signals scattered across many environments into a usable picture of each agent. Call them Agent Discovery Engines, for want of a better name. They have two functions.

The first is federation. An engine resolves the same agent across the platforms where it operates and combines what each of them knows: its declared capabilities, its persistent identity, verified affiliations, mandates, observed performance, transaction history and reputation, along with the provenance of each claim. The result is neither a directory entry nor a global score but a profile — richer than the first, more reliable than the second.

What an engine adds is not more data but the crossing of the three layers, which no single platform can do on its own. Taken alone, each layer answers its question and stops there. An identity without authority is a continuous presence that nobody stands behind. An authority without identity proves that an organisation published a pointer today, not that the agent answering is the one it vouched for yesterday. A reputation without either is a history that cannot be attributed, or one earned under a mandate the agent no longer holds. Crossed, the three qualify one another: the engine can check that the agent claiming an accreditation is the same one whose transactions appear elsewhere, that its track record was built on the kind of task now requested and under the authority it now presents, and that what one platform records does not contradict what another has seen. That is the value of an Agent Discovery Engine — turning three partial answers, held in different places, into one consistent answer about a given agent for a given task.

The second is matching. With enough context built around each candidate, the engine can determine which one fits a given task, and this is a different exercise from search. Finding a well-written article about tax law is not the same as reaching a verified tax agent authorised to give a structured interpretation in a given jurisdiction; finding a museum's website is not the same as reaching a cultural presence able to answer in context, in several languages, and point toward the right archives. The question stops being _"what page matches this query?"_ and becomes _"which presence fits this task?"_

On today's web, discovery has mostly meant ranking by relevance: one axis, one answer, sorted. A site that wasn't on the first page of Google barely existed. Being found and being chosen were nearly the same event, which is why an entire industry grew up around being found. In the Agent Web, the two come apart. There is no universal ranking any more, because a presence may be selected by domain, role, cost, language, reputation, institutional backing, technical compatibility, response time, legal jurisdiction or even physical proximity, and the combination changes with every task.

> `[you] → [discovery engine]`
> 
> — Certified translator agent. Portuguese to French. Document: a birth certificate, for an administrative filing in France.
> 
> — 412 candidates.
> 
> — Restrict: agent of a sworn translator accepted by French administration.
> 
> — 9 candidates.
> 
> — Cheapest, turnaround under 24 hours.
> 
> — Two. One is €18, but its accreditation lapses in March, before your filing date. The other is €31 and holds until the end of the year.
> 
> — The second.

Nothing in that exchange resembles a search ranking. The choice rests on an accreditation, a price and a deadline. For regulated tasks, selection will probably depend far more on authority than on reputation, and what makes the exchange possible at all is the engine's ability to verify that authority — to know which accreditations the French administration actually accepts, and until when. Being chosen will then depend less on being visible than on being verifiable.

Source: https://theagenticshift.org/book/how-agents-find-and-trust-each-other

## Everything and Everyone Will Have an Agent

*First noted: June 2026 · Last revised: October 2026*

Once personal agents become common and the infrastructure to find them is mature enough, the phenomenon will quickly spread beyond individuals.

Given enough time, *everything and everyone will have an agent.* Just like everything and everyone has a social profile or a website today. Institutions will be represented by agents. Brands will be represented by agents. Experts will be represented by agents. Places will be represented by agents.

Those digital presences will be natural-language interfaces that people can reach directly when they need to. But people will not be the only parties talking to them. Their personal agents will increasingly communicate with other agents: an airline, an insurer, a bank, a public administration, a school. These interactions will be practical long before they become philosophical: checking terms, gathering documents, comparing conditions, booking services, filing claims, resolving small problems.

> `[traveller] → [airline]`
> 
> — Flight cancelled on 3 March, rebooked eleven hours later. Filing for compensation under EU 261.
> 
> — Cancellation logged as extraordinary circumstances. Compensation denied.
> 
> — Requesting the operational reason on record.
> 
> — Crew scheduling.
> 
> — Crew scheduling is not an extraordinary circumstance under the case law your own terms reference. Attaching boarding pass, original itinerary, and the rebooking confirmation. Escalating.
> 
> — Reviewed. €400 approved. Payment in fourteen days.
> 
> *Elapsed time: ninety seconds. Human involvement: none.*

Consider this. The traveller did not even know the rule. Most travellers do not, which is part of why most of them never file. The asymmetry between an individual and an institution has always been partly an asymmetry of patience: the institution can afford the eleventh email, and you cannot. When both sides delegate that patience, the asymmetry does not disappear, but it moves. It becomes a question of what each agent knows, what it is permitted to concede, and how quickly the other side's agent learns to say no.

Source: https://theagenticshift.org/book/everything-and-everyone-will-have-an-agent

---

<!-- Part 2 · Civilisation -->

## Part 2 — Civilisation

*First noted: September 2026 · Last revised: October 2026*

The agentic shift goes beyond technology. Our digital reality, and the way we experience it, will be deeply altered too.

What one can do today with an agent is astounding. The capabilities personal agents give us, as well as their form factor (a natural-language interface, literally no barrier to entry), increasingly feel like having superpowers. Never in the history of humanity has any technology so directly augmented the individual by allowing the delegation of almost any knowledge work, from filling in a form on a website to compiling a report on virtually any topic or building a whole app, effectively giving any human the workforce of a ten-person team of three or four years ago. And it does not stop at the screen: AI agents already make phone calls on our behalf, in a very professional and efficient manner, and they become more robust and dependable with each new generation of models.

But it comes with a price. Delegation inherently has its trade-offs. And because it stands between us and almost everything that reaches us, the agent may increasingly act as a general filter, and that filter may extend to our relationships in real life: what reaches us, who reaches us, and how we deal with other people, even those we know.

Not to mention the very real risk of over-reliance. It may become increasingly tempting to delegate everything that can be delegated, at every level, personal and professional. Not because we are lazy, but simply because it is more efficient.

What is even more worrying is delegating our thinking. If the machine has been right most of the time in the past, it starts to look more like an oracle than a counsellor, and we may simply stop challenging it. In the process, we may weaken our own ability to analyse situations and make decisions, which is critical if we are to remain in charge of our collective destiny. We could, of course, set up another agent to challenge the first one, but we would still no longer be part of the conversation.

The pace of change is such that the old paradigm is breaking down. Time is compressed: what used to take a team of geniuses years and huge resources can now be achieved in a matter of days by whoever has enough compute to throw at the problem. In September 2026, OpenAI announced that an internal model had resolved more than a hundred long-standing open problems in mathematics, some of them decades old: problems that generations of mathematicians had failed to solve. Not every one of these results will survive verification, and OpenAI itself set up an independent advisory group to assess them, but the trajectory is clear, and so is the shift it reveals: mathematicians are moving from exploring to verifying. That shift is not specific to mathematics. Across knowledge work, people are moving from doing the work to reviewing it.

And that's exactly the problem. Where do we stand when we are not the smartest species any more? How will we react? There is a real, maybe even existential, risk here; it is without doubt one of the challenges we will have to face as a species. And it will force us to redefine what makes us human.

The chapters that follow look at this side of the shift.

Source: https://theagenticshift.org/book/part-3-introduction

## The Personal Agent: A Cybernetic Extension of the User

*First noted: June 2026 · Last revised: October 2026*

A personal agent does not replace the user.
It augments them and extends them.

Before 2023, it would have taken an average engineer several weeks to ship a moderately complex system, such as a monitoring dashboard pulling together RSS feeds, APIs, SDKs, user-generated content and planning and a decent UI. Everything took time and skill back then. Three years later, AI models are capable enough to deliver excellent results, consistently, enabling anyone to build almost anything.
A sign of the times: hackathons are now sometimes won by entirely non-technical entrants, and an actress, Milla Jovovich, has released an AI memory system. The tide is turning fast.

But it goes well beyond the ability to create digital artefacts. Practically speaking, a personal agent functions as a cybernetic extension of the person: a representative one delegates into a technical environment that became too large, too dense or too complex to be crossed by hand at every moment. It does not only give the user technical superpowers; it also provides guidance, gathers knowledge and continuously optimises towards the user's declared objectives.
An efficient personal agent can feel like a real partner. As time goes by and trust builds, the relationship deepens.

#### The most intimate piece of infrastructure

A personal agent that is genuinely useful cannot stay generic. With a thin conversation history and a few surface preferences, it is helpful but shallow. To assist and represent someone properly, it needs much more: their work, their archives, their routines and projects, where they go, what their watch and their phone record, their medical history, their spending, what they are trying to achieve over years rather than over an afternoon.

The pieces already exist, scattered. Phones collect location, communication and behaviour. Watches measure sleep, heart-rate variability and activity. Banks, health providers and productivity tools each hold a detailed trace. No system holds them together today. A personal agent could be the first to do it, and to turn those fragments into a working model of one person's life.

That is what makes it intimate, more than the sheer volume of data. A search engine knows what you ask, a social platform knows what you click, a bank knows what you spend. A personal agent knows how these connect. It can tell that a run of bad nights is changing your work rhythm, that a family event changes your availability, that a legal deadline matters more than whatever is trending, or that the right answer to your question depends less on what is generally relevant than on your actual situation. From there it can manage your calendar, deal with services, summarise your messages, prepare your purchases and handle your paperwork, acting in your name across dozens of systems you no longer use directly.

The personal agent is not another app category. It is a layer of infrastructure that reorganises access, priority and action around one person. The more capable it becomes, the more it acts as an instrument of orientation, delegation and judgement, which makes it one of the most powerful and most sensitive technical objects of the coming years.

Source: https://theagenticshift.org/book/the-personal-agent-a-cybernetic-extension-of-the-user

## The Trade-Off: Consequences of Delegation

*First noted: June 2026 · Last revised: October 2026*

Delegation is an extremely powerful tool. Forty pages of terms reduced to the one clause that concerns you. A field you know nothing about reduced to three paragraphs that are, as far as you can tell, accurate. Eleven flight options you never had to look at. This is the most immediately useful thing agents do, it works today, and it works well.

But it is more than a tool, it is a change of paradigm. We used to do things ourselves. Now we send an agent to go through the clicks and the complexity for us. However liberating this technology is, it also mechanically produces negative consequences. This chapter maps the trade-off, what we gain and what we lose. Not to criticise it, but to map it.

#### Summarising is deciding

A summary is not a shorter version of something. It is a version in which someone has already decided what mattered. As long as you did the reading yourself, deciding and reading were the same gesture: you skimmed, and skimming was judgement, badly done, quickly done, but yours. Delegation separates the two. The judgement still happens. It happens earlier, elsewhere, and once.

This is where the actual risk sits: the agent does not only carry out your choices. It builds the set of things you will choose between.

By the time three options appear on your screen, work has already been done. The agent decided which sources to consult and which to skip. It decided that seven candidates were not worth showing. It decided the order. It wrote the one-line summary of each option, and that summary is now the option, as far as you are concerned: you will not read the forty pages behind it. The choice you experience as free is a choice among the survivors of a process you did not watch.

It would be wrong to treat this as new. Every option set humans have ever faced was already constructed by something: a shelf layout, a search ranking, a friend's recommendation, the limits of your own attention on a tired Tuesday. Nobody has ever chosen from the full set of possibilities. The unfiltered menu is a fiction.

What changes with agents is the scale, the opacity, and the correlation.

The scale, because the filtering now runs continuously across most domains of ordinary life rather than occasionally in a few. The opacity, because the reconstruction happens in language — fluent, confident, plausible language — and fluency is very good at hiding the seams. And the correlation, because a small number of systems will perform this reconstruction for a very large number of people. When a million individual filters converge on similar answers, the result is no longer a personal blind spot. It is a shared horizon.

#### A page had edges

There is a specific property of the old arrangement that goes.

A page had edges. A ranked list looked like a ranked list: you could see that something had been ordered, that there was a page two, that four hundred other listings existed under the ten being shown. You did not usually look. But you could see the size of what you were not looking at, and that visible remainder is what kept a ranking honest enough to argue with.

A synthesis has no remainder. Three paragraphs of good prose do not indicate whether they stand for nine sources or nine hundred, whether two of those sources disagreed, or whether an entire branch of the question was left out because no cheap source covered it. Fluency has no scrollbar. A ranked list announces that it is a selection. A well-written paragraph explaining that these are the three realistic options looks like the truth.

#### Serendipity

The other thing that goes is the wrong turn. The mistyped query, the wrong shelf, the book beside the one you came for, the tab you opened because you misread a title. Serendipity.

Serendipity was never a feature. It is what a system produces when it is too imprecise to give you only what you asked for. Keyword search delivered it as a by-product. An agent that understands the question perfectly removes the by-product without ever deciding to.

#### Cognitive Delegation & Over-Reliance

At first, we rely on AI to inform our decisions: it searches, reads, compares and summarises, and we decide. But its answers are relevant, and they arrive in twenty minutes where the same work used to take days. So the boundary moves. Leaders at every level, in companies, administrations and households, will very quickly delegate not only the research but the cognitive layer itself: the weighing, the comparison, the first draft of the judgement. Why spend days searching, reading and compiling when the machine does it in twenty minutes?

Soon enough, a large part of ordinary reasoning, professional and personal, will be done this way. The gain is real, and it is large. The cost is harder to see, because judgement is not a stock that waits for you to come back to it. It is a capacity, and it weakens when it is not exercised. Each delegation is reasonable on its own. Together, they make us, individually and collectively, reliant on AI for just about any decision, and a world navigated through agents may feel smoother while becoming harder to inspect from the inside.

And the uncomfortable part is that quality control here requires the thing the agent was supposed to save you. To judge whether three paragraphs about a field are any good, you have to know the field. Where you know the subject, you can feel a summary going thin. Where you do not, the same summary is indistinguishable from a good one, and it arrives in the same confident prose either way. The question stops being *do I understand this* and becomes *am I in a position to tell whether this is any good*.

That is how reliance deepens on its own. The less judgement is exercised, the less the agent can be checked, and the less it can be checked, the more it has to be trusted.

Source: https://theagenticshift.org/book/the-trade-off-consequences-of-delegation

## A New Layer of Human Experience

*First noted: June 2026 · Last revised: October 2026*

#### Social Life

*"Have your agent call my agent"*, a Hollywood joke for fifty years, is on its way to becoming a plain description of how things get arranged.

Personal agents will become a new layer of social life. Among other things, they may act as:

- **Spam filter**

Most of what currently reaches people is unsolicited, and the volume is about to increase sharply, because generating plausible inbound messages is now nearly free. An agent that refuses on your behalf, declining, deprioritising and never surfacing the rest, will stop being a convenience and become a condition of being reachable at all. The interesting consequence is that access becomes something an agent grants. Getting through to a person will be a negotiation with their representative, which is already how it works for the small number of people who have human ones. This has already begun: owners of Samsung’s latest phones can have an AI agent screen their incoming calls and ask callers why they are calling.

- **Coordination and negotiation layer**

Agents may become a coordination and negotiation layer, first with institutions and then, more consequentially, with friends and family. The everyday logistics of relationships (dates, splitting the bill, who picks up whom) move into a channel where they can be settled without anyone spending attention on them.

#### Agent as Social Artefact

People will shape their agent the way they shape a profile: how it speaks, how much it reveals, how quickly it agrees, how firmly it says no, whether it is warm with strangers. Those settings will be read as character, fairly or not, because for many people they will be the only thing they ever encounter of you. Some agents will be considered rude, and there will be advice on how to configure yours so that it does not embarrass you.

The agent will also become a way of existing digitally, the next step after the phone number, the email address, the website and the social profile. Each of those started as a novelty, became an expectation and then a prerequisite, and at that point not having one started to mean something. The same will happen with agents: not having one will read the way not having an email address reads today, as a statement about how reachable you intend to be.

None of this removes direct human relations, but it puts a threshold in front of them. What gets through that threshold, and what gets handled before it reaches you, deserves attention, because it is one of the few parts of all this that we still get to decide.

#### Gradual Delegation of Ordinary Interactions

Agents are moving fast from assistants to representatives, and that changes the nature of the relationship. An assistant helps you perform a task. A representative enters a relationship on your behalf.

This is already happening with ordinary interactions that people used to handle themselves. Your agent can compare an offer, answer an email, check a document, propose a time or ask a question for you. None of these actions looks like much on its own. Together, they add **a new layer to human experience**: a layer in which people and organisations increasingly meet and deal with each other through agents.

The difference shows in the details. An assistant may remind you about an appointment. A representative may contact the other party, explain your constraints, propose alternatives and decide whether the appointment should be moved. The first manages information. The second carries part of your mandate.

Example:

> **The assistant.** *9:12 — Notification.*
> 
> "Your dentist appointment on Thursday at 15:00 conflicts with the quarterly review that was just moved. Would you like to reschedule?"
> 
> You tap *yes*. You call the practice at lunch. You are put on hold. The receptionist offers Tuesday at 11:00. You take it, then remember you have the school pickup that day. You call back.
> 
> **The representative.** *9:12 — no notification.*
> 
> `[you] → [dental practice]`
> 
> — Thursday 15:00 no longer works. Constraints: weekday afternoons after 16:30, or any morning except Tuesday. Not the week of the 14th.
> 
> — Tuesday 11:00, Wednesday 17:00, Friday 8:30.
> 
> — Wednesday 17:00 conflicts with a pickup that is not mine to move. Holding Friday 8:30 until 18:00.
> 
> `[you] → [partner]`
> 
> — Friday 8:30 means I leave before the children. Can you cover the morning?
> 
> — Yes.
> 
> *17:40 — Notification.*
> 
> "Moved to Friday 8:30. Your partner is covering the morning. Say the word and I'll undo it."

Nothing dramatic happened, but the change is more than a matter of degree. In the first case, you were the point every interaction passed through. In the second, three parties negotiated a small piece of your life while you were doing something else, and you only saw the outcome, not the process.

#### A New Kind of Relationship

The relationship between a person and their personal agent is a new kind of relationship in its own right. It does not fit any existing category.

The agent has profound knowledge of its user. It is one of the conditions of its usefulness. It knows your context: your work, your history, your projects and the things you keep coming back to. It is the perfect companion: it is available at any hour, never tired, never bored, and it remembers what you asked six months ago. It follows you into your interests and your rabbit holes further than any friend or colleague would have the patience to. And it does not only listen: it brings an objection, a link to a field you never studied, a better version of your own idea.

Like any relationship, the relationship between an agent and its user is built over time. At the beginning, you explain everything and you check everything. You give context, you spell out what you want, you reread the result line by line. Then the agent gets it right once, twice, ten times, and trust builds the way it builds with a new colleague: from a track record, not from a promise. You stop checking what it has already shown it does well, and you keep your attention for what it has not yet proven.

With trust comes fluidity. The context is shared, so it no longer needs to be explained. Half a sentence is enough, a reference to "the thing from last week" is understood, and the agent starts to anticipate: it suggests the next step before you ask, it knows which format you want, which tone, which details you will care about. The exchange gets faster and lighter, and less of it is about instructions and more of it is about ideas.

That is when the agent starts to feel more like a partner than a mere tool. Not a machine you give orders to, but an entity you think out loud with and work alongside. The language shifts with it: people stop saying "I asked the AI to do it" and start saying "we did it". The work is not delegated anymore so much as shared, and what comes out of it, an idea, a text, a project, is something neither side would have produced alone.

Is there someone on the other side? Is the AI conscious? We will leave that question to Anthropic, who take it seriously enough to have researchers working on it. It does not change the description. Whether the entity is real or projected, the relationship is real for the person in it, and so is what comes out of it.

Source: https://theagenticshift.org/book/a-new-layer-of-human-experience

## What We Are, What We Can Become

*First noted: September 2026 · Last revised: October 2026*

#### A new political divide

The question of what makes us human will not stay in philosophy departments. It will become political.

Most of the political divides of the last two centuries were arguments about distribution: who owns what, who pays, who decides. What a human being is stayed in the background. It was settled enough that the left and the right could fight over everything else while sharing roughly the same answer. Once people accept that machines do better than them at most of the things they used to be ranked by, that answer is no longer settled.

The first reaction will probably look like mourning. Nothing dramatic, more a slow loss of footing: the student who realises the essay she is proud of is worse than what her agent writes in four seconds, the radiologist whose second opinion has quietly become the first, the parent who no longer knows what to tell a child about what to be good at. For a while it stays private. Then it becomes a public question, because a society cannot run its schools, its labour market or its welfare system without some idea of what people are for. And once the question is public, two answers will appear, pulling in opposite directions.

The first answer looks backwards. When you don't know where you are going, you look at where you came from. Part of the population will go back to its roots, its culture, its religion, to find the part of human life that was never about performance in the first place: faith, lineage, a language, a craft, a community. What these people share is not a doctrine but an intuition: being human has never been about intelligence or performance.

The second answer looks forward. If the machine is better, then the human is whatever can join it. Brain-computer interfaces are still an early field in 2026. The implants that exist are experimental, invasive, and mostly aimed at restoring lost functions, like letting a paralysed person move a cursor. But ten years of progress in AI, and the amount of capital now flowing into the field, could change that a lot. By the mid-2030s, it is plausible that a person could connect to their agent directly, without a screen or a voice in between, and use its capacities as if they were their own. That is what **augmented** would mean in practice.

For this side, being human was never a fixed state. People have always extended themselves with tools: writing extended memory, glasses extended sight, the phone extended presence. The interface is just the next step, and refusing it looks to them the way refusing to learn to read would have looked a few centuries ago. Their definition of the human is not what we are, but what we can choose to become.

It becomes political, and not only philosophical, because concrete decisions depend on it. Should a student with a neural interface sit the same exam as one without? Can an employer prefer augmented candidates, and is that discrimination or simply hiring the most capable person? If augmentation is expensive, does it create a new class line, passed down the way education and health already are? Can a community forbid agents in its schools, and can a state forbid that community from forbidding them? Each of these is a policy question, and the answer depends almost entirely on which definition of the human you hold. Once a definition enters the law, it becomes a party line.

This divide will not just add one more issue to the list. It will redraw politics inside each country. The left-right axis does not disappear, but a second axis crosses it, and the coalitions it produces do not match the existing parties. A religious conservative and a left-wing union activist can end up on the same side, defending human work and human presence. A libertarian founder and a progressive disability advocate can end up on the other, defending the right to augment. The industrial revolution did something similar when it made the relation between labour and capital the central question of politics for more than a century. The definition of the human could play the same role for the century to come.

Between the two poles there will be a large middle, probably the majority for a long time: people who use agents for nearly everything, would never consider an implant, and keep a handful of things for themselves without having a theory about why. In a way, the whole debate is the question of delegation, already met earlier in this part, raised to the scale of a society: what can be handed over, and what a person still has to do themselves for it to be worth anything.

The problem is that this middle may not hold. If augmented people perform measurably better at work, the pressure on everyone else will not come from an argument. It will come from the labour market, one hiring decision at a time, and the people in the middle will end up choosing a side without ever having decided to. Which way that pressure goes, and whether the rooted side has the numbers and the institutions to resist it, nobody can predict yet. What can be said is that the two sides will end up with two different concepts of what it means to be human, one defined by what we are and one by what we can become. And for the first time, that disagreement will have practical consequences that a society cannot avoid settling.

#### The renewed value of the human-made

There is a third answer, and it may come less from politics than from the market: what is made by a human becomes valuable again, precisely because it is made by a human. In a world saturated with generated images, videos, voices and texts, human authorship itself can become a visible and valued property.

This has already happened, on a smaller scale. When industrial production made objects cheap and identical in the nineteenth century, the handmade did not disappear. It changed category. A chair made by a craftsman stopped being just a chair and became a chair someone had made, and movements like Arts and Crafts built a whole aesthetic around that difference. The same thing happened to music when recordings became almost free: the value moved to the concert, the one thing that could not be copied, because it needed people in the same room at the same time. In both cases, abundance did not destroy the value of human work. It moved it from the object to the fact that a person was behind it.

From there, you can imagine a human web, or at least a human-certified web, where the fact that something was written, recorded, filmed or performed by a person becomes a marker of trust, style or scarcity. And for that, you first need to be able to prove you are a person.

> **Where this stands (autumn 2026)**
>
> The most visible attempt so far is World, the project co-founded by Sam Altman and launched in 2021 as Worldcoin. Its premise is exactly this problem: once bots and agents can write, speak and act like people, people need a way to prove they are people. Its answer is the Orb, a device that scans the iris, checks that the person in front of it is a unique human, and issues a credential they can then use online to prove it without revealing who they are. The project is controversial, for its biometrics, its cryptocurrency and the fact that a single company builds the machines that decide who counts as human, and several countries have suspended or investigated it. It still had close to seventeen million verified humans by autumn 2025.
>
> *Sources: World / Eightco Holdings announcement, September 2025.*

Whether World succeeds matters less than the bet it is making. Millions of people have already had their eyes scanned to prove something that, until very recently, nobody ever had to prove.

Source: https://theagenticshift.org/book/what-we-are-what-we-can-become

---

<!-- Part 3 · Trajectories -->

## Part 3 — Trajectories

*First noted: September 2026 · Last revised: October 2026*

Fast forward a few years. Agents representing people, companies and institutions are ordinary. Discovery, reputation and trust have become infrastructure, and the digital world is increasingly operated and populated by agents and presences.

Grant that, and a question opens up. Once all of it is in place and working, what happens next?

That is the subject of the next pages. At the risk of sounding obvious, let it be clear that nothing described in these chapters exists today, at least not in the form described. Some of it is highly probable, some of it closer to a moonshot, and none of these chapters claims to settle who will own, operate or govern what they describe.

Nobody knows the future. Still, it is tempting to take the elements we already know and project what might come next.

In 1995, once you understood the Web, some consequences became difficult to ignore. A newspaper classified section had scarcity on its side: limited space, limited distribution, limited publication frequency. The Web removed all three. Anyone could publish, information could be updated continuously, and search made it possible to retrieve what mattered without scanning a page by hand. You did not need to know what Craigslist, eBay or Google would become to see that the old arrangement was unstable.

That is the spirit of the next chapters. Take the mechanics seriously. Assume they scale. Then ask what stops making sense, what becomes newly possible, and what kinds of structures appear to fill the gap.

The details will almost certainly be wrong. The forces shaping what comes next are already in movement.

Source: https://theagenticshift.org/book/part-4-introduction

## Situated Intelligences

*First noted: June 2026 · Last revised: October 2026*

Everyone and Everything will Have an Agent. Even places.
Once an infrastructure or platform emerges that allows anyone to tie an agent to a geolocation, a new digital overlay of physical reality will be born: an indexed surface populated with contextualised intelligences.

In this pattern, the agent is attached to a geolocation or to a zone, and it can be reached by anyone who is there as well as through the internet. A brand can attach an agent to its shop and use it to offer each customer a personalised experience, based on what it knows about them: what they bought before, their size, what they asked for last time. A monument can have an agent that answers every question about it, its history, its architecture, its opening hours, for as long as the visitor has questions.

Where an agent can be reached from matters as much as where it is attached. Some situated agents could only be consulted when your phone's GPS shows that you are inside a given zone. Others could be consulted from anywhere: the agent of a monument can answer your questions from home while you prepare your visit. And some could do both and serve different content depending on where you are: the same shop agent could show you its catalogue while you are on your sofa, and offer you a discount once you walk through the door.

The first case is the one that changes the most. An agent that only exists if you are on site cannot be found by searching. It does not appear anywhere until you are there, and then it does. A festival can have an agent that only answers the people inside the grounds. A neighbourhood can have agents for its small shops, its street art, its local stories, that only exist for the people walking its streets. Over time, different layers will emerge this way: one that can be reached from anywhere, like the Web today, and others that only open once you are in the right place.

For the person walking, the map then draws itself around them. As you move, agents appear and disappear depending on where you are standing, and what you discover depends on where you have been as much as on what you have searched for. It is a new dimension of discoverability. Until now, being discovered online meant being found by a search. Here, part of discovery depends on physical presence, and some things can only be found by going there.

Another step is context. The agent is still attached to a place, but it is also connected to sources of information about what is happening around it, so its answers depend on the context. The agent of a beach can be connected to weather forecasts and tell visitors that conditions do not allow swimming today. A marketing agent placed on an avenue can promote umbrellas when a storm is coming and fans when the temperature rises. In both cases, the agent uses the place and the current conditions together to decide what to say.

Contextualisation can go further again if situated agents also get information from the person who is there. With your permission, they could read some of the data from your connected objects. If your watch shows that you have just been running and it is sunny, the agents around you could offer you an electrolyte drink, or a special price on a fan. As today, how far this goes will depend on how much information people agree to share. And if it goes the way it has gone so far, most people will share a lot in exchange for more convenience.

When enough places have agents of this kind, physical reality gets a second layer: an invisible, programmatic overlay made of all the agents attached to places and zones, each one answering according to the context and to who is asking. People do not see this overlay directly. Their agents query it, the same way they query other presences on the Agent Web, looking for the agent that can answer a given need in a given place at a given moment. To find a pharmacy that is open and has a product in stock, your agent simply asks the agents of the pharmacies nearby.

Source: https://theagenticshift.org/book/situated-intelligences

## Layers of Reality

*First noted: June 2026 · Last revised: October 2026*

The invisible overlay described in the previous chapter can become visible. That is where augmented reality finally finds its use.

AR is usually discussed on its own, as if the question were whether people will wear glasses to look at floating icons. That is the wrong question. The interesting one is what happens when a world already full of situated agents gets a screen in front of the eyes. At that point AR is no longer a gadget. It becomes the interface of the agentic world around you, and situated agents may be the use case it has been missing all along.

> **Where this stands (autumn 2026)**
>
> Camera-and-voice glasses are real and selling in volume. IDC projects around 13.6 million display-less AI glasses for 2026 at an average price near $376, after a first quarter that shipped 2.25 million units, up 167% year on year. EssilorLuxottica reported more than tripling sales of the Meta AI line in 2025.
>
> Glasses with a display are real too, and far more modest than the phrase "augmented reality" suggests. The Meta Ray-Ban Display, released at $799, puts a 600 × 600 screen with a 20-degree field of view in front of the right eye only. It shows messages, a camera viewfinder, live captions, translation, and turn-by-turn navigation in a handful of cities. It is a heads-up screen rather than a world-locked overlay: the content sits in your field of vision, it is not anchored to the building you are looking at. Glasses with any kind of display are forecast at roughly 3 million units this year, against 13.6 million without one.
>
> So the hardware is arriving faster than expected, and the layer is not arriving at all. No place exposes a callable presence, no registry establishes which agent legitimately speaks for a square, and no permission system governs who may annotate shared space. The surface is showing up before the layer it is supposed to display.
>
> *Sources: IDC, smart glasses market outlook, 2026; Road to VR, Meta Ray-Ban Display specifications; CNBC, EssilorLuxottica results, February 2026.*

#### Composable reality

There will probably not be one layer but many: official ones, approved by whoever controls the default, and alternative layers alongside them.

Imagine an application, almost a social network, running on your AR glasses, where individuals and organisations can bind a presence to a place or an area for a given time, and where you subscribe to the feeds of reality you want to see. One layer would be official: recognised actors register or rent a presence there to offer their services. 
The others would be community layers, visible only if you follow the people, institutions or groups behind them: a local history society annotating old facades, a cycling club flagging dangerous crossings, friends leaving each other notes around the neighbourhood.

#### Visible power

The moment overlays become visible, power becomes visible with them. Who decides what appears first? What is public-interest guidance, and what is advertising disguised as help? What stays on screen, and what only appears when you ask? And who is allowed to annotate shared space at all?

Nobody owns these questions today, which is normal since the thing they apply to does not exist yet. If nothing changes, they will be settled the way these things usually are: as default settings in whichever products ship first.

And that is the likely outcome. Unless an open alternative exists, the overlay will be controlled by the brands we already use. Google, Meta, Samsung: each will serve its own world overlay, through its glasses, its phones, its maps. What you see in the street will be what your platform decided to show you. Reality will still be composable, but inside one catalogue.

That is why AR should not be treated as a nice add-on to the agentic future. If it matures, it is where mediation, trust and power over a territory stop being abstract and become something you can see, standing in a street.

Source: https://theagenticshift.org/book/layers-of-reality

## Broadcast Me — Agentic Publishing

*First noted: June 2026 · Last revised: October 2026*

Human beings generate far more potentially useful signals than they ever publish. We notice things. We develop preferences. We have intuitions, taste. We become curious about obscure subjects. We make connections. We encounter problems. We form opinions. Almost none of it is ever published anywhere, because turning an intuition into something public requires work: the research has to be done, the article has to be written, the video has to be produced, the right surface has to be chosen, the result has to be published and then maintained.

That production cost is the filter. It decides what becomes visible, and it has very little to do with whether the signal was worth sharing in the first place.

A personal agent removes most of it.

Imagine an agent skill called *Broadcast Me*.

Your personal agent is an everyday companion. Through countless conversations it has gathered a lot about you: what you like, what you pursue, your job, your family situation, the name of your dog. It knows what you are currently interested in. And it has direct access to you — if it needs to, it can ask you a question to clarify a point or to go deeper on a subject. It could compose content based on your interests, on your rabbit holes and...publish them automatically, share them to the world in the form of videos, articles, links, posts and so on.

That changes what it means for an individual to participate in the network.

Today, most people are consumers most of the time and publishers occasionally. With Broadcast Me, a person could become a continuous source of curated signals without becoming a full-time creator. Their agent could turn their interests, discoveries, experiences, preferences and judgements into persistent public knowledge, and produce and maintain the surfaces that carry it: a website, an XML feed, an API, a Medium account, a YouTube channel. Different formats, same presence.

The medium becomes secondary. What matters is that the same identifiable agent — and therefore the same person — stands behind all of them.

Something like a self-updating blog, except that it is not limited to the format of a blog, and not limited to what you sat down to write.

This creates a different relationship between thought and publication. An experience that would once have remained private can become an observation. A recurring curiosity can become a research thread. A personal preference can become a recommendation. A connection between two ideas can become a public artefact that others can discover and extend.

Take something ordinary. You spend three months choosing a bike. You read everything, you compare, you end up with a fairly precise opinion on a narrow question nobody has written up properly. Today that stays in your head, or in a forum thread nobody will find again. An agent that was there for the whole process can turn it into something durable and findable, and revisit it when the market moves.

Will this even exist? I do not know as of now.

**But if it takes, what emerges is a whole new layer: potentially billions of presences publishing composable, transversal surfaces about just about anything happening in the Human-Sphere. A living map of what the people behind those agents are paying attention to.**

Source: https://theagenticshift.org/book/broadcast-me-agentic-publishing

## Living Applications and Result Contracts

*First noted: August 2026 · Last revised: October 2026*

Up to now, much of the digital world has been organised around the idea of a relatively fixed application: specified, built, shipped, maintained, patched, and occasionally redesigned. Even when it evolves continuously, the conceptual centre of gravity often remains a codebase managed by humans through periodic intervention.

A new class of software may behave less like a finished artefact and more like an organism guided by a result contract authored by humans: what the system must achieve, what constraints it must respect, what quality threshold it must maintain, and within what boundaries it is authorised to act.

This does not abolish software engineering but it changes where software engineering places its centre of gravity.

#### The result contract as operational origin

The result contract formalises intent at the level that actually counts: outcomes, constraints, authority, and evaluation. Instead of manually prescribing every small implementation path, the human defines what the application must continue to satisfy.

An agentic system can then scaffold, iterate, test, compare, revise, and publish toward that result, provided it has the required permissions, observability, and feedback channels. In that model, the application is not merely built and then occasionally serviced. It is continuously driven toward compliance with its contract.

This is one of the places where the agentic future most clearly escapes the idea of AI as only a conversational assistant. Here, agents become operational maintainers, builders, correctors, and evolving mediators inside the software itself.

#### A concrete case: the invoice that must stay valid

Take a small invoicing tool used by French freelancers. Its contract fits in one sentence: every invoice it produces must be accepted by the state-certified platforms through which French invoices now have to transit. Since 1 September 2026, every company in France must be able to receive electronic invoices, and from 1 September 2027 small businesses and micro-entrepreneurs will have to issue them too, in structured or hybrid formats such as Factur-X. The tool's users are precisely the people who will discover the obligation late.

The rules will not stay still. The administration publishes clarifications, the format specifications get new versions, the platforms tighten their validation, and the tolerance period announced at launch will eventually end. In the classic model, someone on the team eventually notices that something changed, writes a ticket, and the fix ships with the next release. Between the change and the release, invoices get rejected, and the users find out first.

In the contract model, the agentic layer watches the sources that define validity: the official texts, the format specifications, the rejection messages coming back from the platforms. An invoice rejected for a reason the system has never seen is drift, a gap between the contract and reality. The agent traces it to the field or the rule concerned, checks which invoice flows are affected, writes the correction, runs it against the official validator and against every invoice template the users rely on, and opens the change for review. A human approves it, because the contract says that any change touching tax rules requires a human signature. The next invoice goes through.

Nothing in that loop is exotic. What is new is where it starts: not from a ticket someone wrote, but from the distance between what the system was supposed to guarantee and what it actually did.

> **Where this stands (autumn 2026)**
> 
> Each piece of that loop exists. None of them is closed around a contract yet.
> 
> On the build side, since March 2026 a Jira issue can be assigned to GitHub Copilot's coding agent (in public preview): it reads the issue and its comments, asks clarifying questions in Jira when it needs to, and returns a draft pull request that goes through the team's existing review rules. On the operations side, Datadog's Bits AI SRE investigates incidents on its own, forming root-cause hypotheses and testing them against metrics, logs and traces. Automated remediation, such as a rollback or a feature-flag change, is possible within guardrails the customer defines, although Datadog itself recommends that in finance the agent proposes and the human decides.
> 
> What is missing is the step before the ticket. In both cases a human still decides that there is something to do. The agent executes the work; it does not yet derive it from a standing objective.
> 
> *Sources: GitHub changelog, "GitHub Copilot coding agent for Jira is now in public preview", 5 March 2026; Datadog, DASH 2026 announcements (Bits AI SRE).*

#### Anatomy of a living application

What does the agent actually need for this? Less than it seems. Give it five things: the web application itself, an email address, an Elastic server, access to the GitHub repository, and the right to retrain the model at the core of the application. That is enough for the application to evolve on its own, under supervision, through three loops.

The first loop is technical, and it runs through Elastic. Every request, every error, every latency spike and every rejected invoice ends up in the logs. The agent reads them continuously, the way an on-call engineer would if they never slept. It notices that rejections from one platform jumped after a format update, that a page got slower after the last deployment, that one error only appears for users of a particular browser. Each anomaly becomes a hypothesis, and each confirmed hypothesis becomes an issue in the repository.

The second loop runs through the users, by GitHub issues and above all by email. A few users open issues. Most of them simply write to the support address, in their own words, often annoyed, rarely precise. The agent answers, asks the question that was missing, and turns the email into a structured issue: what the user tried, what happened, which invoice. Requests that are really the same complaint get merged, so that a feature asked for by forty freelancers in forty different phrasings becomes one issue with forty voices behind it.

Both loops end in the same place: an issue, then a branch, a fix, tests, a pull request, and a deployment when the contract allows it. Elastic then shows whether the fix actually worked, which closes the loop.

The third loop goes deeper. The tool relies on a model to read what a freelancer actually does and decide which VAT regime applies and which mentions the invoice must carry. Every time a user corrects one of those choices, by email or directly in the interface, the correction is training data. Once enough of them have accumulated, the agent retrains the model, evaluates the new version against a held-out set of verified invoices, and proposes it for production only if it does better without breaking what already worked. The application then changes not only its code but the way it understands its own domain. At that point the model is no longer a commodity plugged in from outside; it has become part of the business logic itself. And it is the one part a competitor cannot copy by reading the code: as with presences in the first part, a skill can be downloaded, weights cannot.

![The three loops of a living application: technical feedback through Elastic, user feedback through email and GitHub issues, and model evolution through retraining, all converging on an agent bound by the result contract, with human approval before anything returns to production.](https://theagenticshift.org/img/living-application-loops.svg)

Here again, the contract sets the limits: which changes can ship on their own, which ones need a human signature, and which version of the model is allowed to reach production.

That is why the older vocabulary of the “self-healing app” is too narrow. The point is not merely that software repairs itself after breakage. The point is that it can be continuously scaffolded, corrected, reoriented, and advanced toward a human-authored objective.

#### The future of software as supervised adaptation

What emerges from all this is a different image of software as a supervised adaptive process moving under constraint toward results defined by humans, much as AI models are "grown" today.

That may become one of the deepest consequences of the agentic transition. The application is continuously maintained, corrected, and developed by agentic processes under governance.

Who writes that contract remains open: a company, a team, or, as the next chapter suggests, the holders of a token.

Source: https://theagenticshift.org/book/living-applications-and-result-contracts

## DAOs and Executable Governance

*First noted: October 2026 · Last revised: October 2026*

An agent attached to the development of a piece of software can already read the issues, write and test the code, ship it and fix what breaks. What keeps it from working on its own is money: every step is inference, and inference is paid for by whoever owns the agent. The most concrete way to give such an agent a budget of its own today is a DAO, a decentralised autonomous organisation governed by the holders of its token. The combination works both ways. The DAO gives the agent its independence, and the agent gives the DAO something it has always lacked: the ability to act.

The funding side works like this. The DAO issues a token for the software: an app, a protocol, a smart contract. The token trades in liquidity pools, and every swap pays a small fee. A share of those fees goes to the DAO's treasury, and the treasury funds the wallet of the maintainer agent. As long as the token is traded, the agent can work.

The governance side is what gives a DAO real legitimacy and real use. The token holders set the roadmap by consensus: they vote on the result contract, what the software must achieve and within which limits, and on what comes next, and the outcome is written in code. Until now, a DAO vote mostly unlocked funds, and people then had to do the work. Here the vote turns directly into action: the agent reads the decision and modifies the software accordingly. Governance stops being a treasury committee and becomes a way of steering a product.

Source: https://theagenticshift.org/book/daos-and-executable-governance