This report is written under Chatham House Rules. No statements are attributed to individuals.

In late March 2026, David Caswell and Shuwei Fang convened roughly forty people (publishers, technologists, investors, journalists, researchers, builders) for a two-day working summit called Signals at Scale, hosted at the Shorenstein Center at Harvard Kennedy School. It was the third in a line of such summits (the Bay Area and New York were visited first). The goal was explicit and ambitious: stop iterating on the margins of what the information economy looks like today, and start thinking seriously about entirely new architectures for how information is produced, distributed, valued, and consumed in a world where AI has become the mediating layer through which most people encounter most of what they know.

The operating assumption, established in the keynote and reinforced throughout, was that we are in an efficiency phase: AI applied to existing tasks, within existing workflows, to produce existing information products. And that this phase, however profitable in the short term, is a dead end. The fundamental structure of the information ecosystem is changing. The question is what the new thing looks like, and whether the people in the room could help steer it toward outcomes that are societally positive rather than merely commercially expedient.

I came away with one conviction that was stronger than any I arrived with: the most consequential question in the emerging information ecosystem is about distribution. Specifically, it is about the architecture of personal AI agents, what they negotiate, on whose behalf, and under what terms. That architecture will determine whether the new ecosystem empowers people or simply replaces one set of gatekeepers with another. And within that question, the deepest issue is the integrity of intention itself: whether agents will faithfully represent what their principals actually need, or whether intention will become the new attention, a signal to be captured, distorted, and monetised. This report covers the summit’s intellectual framework, the working sessions, and the distribution thesis that crystallised for me at the investor table. It is written for colleagues who were not present.

The Summit’s Intellectual Framework

Situational Awareness: The Capability Trajectory

Caswell and Fang opened with a calibration exercise, forcing the room to confront the actual trajectory of AI capability. The demonstration was simple: the same prompt given to GPT-2 in 2019, GPT-4 in 2023, and o3 in early 2025. The progression from incoherent babble to competent analysis to a PhD-level 25-page treatise made the abstract concrete. A further comparison placed the latest model’s capabilities against human benchmarks, and the crossover point (the moment machine performance surpasses the median human across digitally accessible domains) was presented as imminent.

The harder point was what comes next. The benchmarks are real and accelerating. The infrastructure investment is staggering: over five trillion dollars committed through 2030, with data centres scaling to campus-sized installations. The frontier labs are publicly stating that fully automated AI research may be plausible by early 2027. Whether you take that timeline literally or discount it by half, the structural argument holds: we are adapting to something that is becoming an autonomous economic actor. The keynote also introduced the concept of ‘jaggedness,’ the uneven capability profile that makes AI hard to reason about. In 2023 it was a fun toy. By 2024 it was astonishingly intelligent in some domains and bafflingly stupid in others. By 2025 and 2026, the jaggedness is smoothing out. The gaps that allowed people to dismiss AI as unreliable are closing, and the awareness gap between those who work with frontier models daily and those who formed their impressions two years ago is widening into a strategic liability.

Legacy publishers’ instinct is to find reasons why the disruption won’t be as bad as feared. Every objection (it’s hype, it’s a bubble, the scaling laws are plateauing, the lawsuits will constrain them) was acknowledged as partially true and entirely beside the point. The direction of travel is clear, and building strategy around the hope that it slows down is not a strategy.

Four Paradigm Shifts

The conceptual core of the summit was organised around four emergent shifts in the information ecosystem, each representing a move from a legacy assumption to a new structural reality:

– Scarcity to Abundance. The gathering, processing, distribution, and experience of information are all becoming abundant. Human attention remains finite, but agentic attention (AI systems consuming and acting on information on our behalf) is effectively unlimited. The constraint is no longer the supply of information but the architecture for making it useful.

– Human Audience to Machine Audience. The consumer of most information is no longer a person reading an article. It is an AI system ingesting, disaggregating, and remixing content to serve a query or fulfil a task. The summit framed this as a shift from B2C to B2A2C, and ultimately to B2A2A2A2A...C, a chain of agentic intermediaries between producer and person. The agentic internet is nearly operational.

– Artifacts to Liquid Information. The fundamental unit of the information economy has been the artifact: the article, the report, the broadcast segment. That unit is dissolving. Information is becoming fluid, anything-to-anything, separating semantics from syntax, flowing through systems rather than sitting in containers. The metaphor offered was buckets versus pipes. If information is liquid, the container (the publication, the brand, the format) loses its structural role.

– Attention to Intention. The business model of the legacy information economy was built on capturing attention: what can I get you to look at? The emerging model is organised around fulfilling intention: what do you actually need? This is the deepest of the four shifts, because it changes the business model and the moral logic of the entire system. I will return to this at length, because intention is both the most valuable signal in the new ecosystem and the most vulnerable to manipulation.

Listen to a NotebookLM-generated podcast about this field report here:

The Barbell and the Bullish Case

One of the more useful economic frameworks presented was the barbell effect: as information becomes liquid, the middle of the market hollows out. Pure commodity (low cost, near-zero margins, scale moat) and pure premium (high cost, high margins, differentiation moat) both survive. Everything in between gets crushed. For publishers operating in the middle of this distribution, producing competent general-interest content that is neither the cheapest nor the most differentiated, the barbell is an existential diagnosis. But the summit was explicitly oriented toward a positive case.

The argument: we may be witnessing the largest expansion of the market for information in 500 years. Supply expands because production costs collapse and deadweight loss (information that exists but isn’t usable) gets unlocked. Demand expands because consumption becomes more efficient, because new agentic consumers emerge, and because latent demand that could not be expressed in the old system gets surfaced. The total addressable market for information was presented on a historical timeline: writing, the printing press, radio, the internet, mobile, and now AI, each expanding the market by orders of magnitude.

The critical insight: information overload is a red herring. We are overloaded by the quantity and inefficiency of legacy information artifacts. We are overloaded by the constraints of accessing and understanding information via those artifacts. We are overloaded by our inability to delegate the consumption of information. We are not overloaded by information itself.

What Breaks When Your Customer Is an AI?

I spent a significant portion of the summit in a working group focused on the needs of information producers. The exercise was to identify first-principles requirements: the most elementary needs that must be met for producers to function in the new ecosystem. The initial list was familiar: purpose, sustainability, trust, relevance, recognition, safety, creativity, curiosity. A useful taxonomy of layered output emerged (information as raw facts, insight as interpretation, intelligence as actionable synthesis for complex decisions) with the observation that the business models, talent requirements, and defensibility differ significantly across layers.

The Blindspots: Who Inherits Their Exploitation or Mitigation?

I offered what I think is a useful way of looking at the industry I am in: traditional publishers are not in the business of producing information. They are in the business of exploiting or mitigating blindspots in human cognition and information gathering. This describes the full spectrum of what publishers do. At the mitigating end: investigative journalism reveals what powerful actors have hidden. Foreign correspondence shows you what is happening in places you cannot go. Data journalism makes visible the patterns in complex systems that are invisible to unaided perception. Science journalism translates knowledge between expert and public communities. All of these compensate for inherent limitations in what any individual human can know, perceive, or process. At the exploiting end: clickbait leverages the gap between what you think you want to know and what is actually useful to you. Sensationalism exploits the cognitive bias toward threat and novelty. Filter bubbles exploit the human tendency to seek confirmation rather than challenge. Advertising-funded content exploits the misalignment between what holds your attention and what serves your interests.

Most publishing sits somewhere between these poles, and the same organisation often does both. The structural point is: if publishing’s function is to interact with cognitive blindspots, then the question in an AI-mediated ecosystem is who inherits that function. In the agent model, the answer is the agent. And that makes the agent’s integrity, whether it mitigates or exploits its principal’s blindspots, the most consequential design decision in the new information economy.

Old Ways of Thinking

With the blindspots thesis on the table, someone made the observation that sharpened the rest of the session: this list of producer needs could have been written in 1996. The needs haven’t changed, but the structural reality has. In an AI-mediated ecosystem, the producer’s customer is the AI intermediary, not the end user. And that single structural fact breaks nearly everything on the list.

– Trust shifts from audience trust to machine trust: whether AI systems treat your content as reliable and worth surfacing. This is partly technical (is your content well-structured and attributable?) and partly contractual (do you have a licensing relationship that gives you leverage?). The reputational signals that humans use to evaluate trustworthiness, brand recognition, editorial voice, track record, may not translate into the signals that AI systems use to weight sources.

– Feedback loops break entirely. You no longer know which of your content reached which users, in what form, or with what adjacent content. The feedback is filtered through an intermediary whose optimisation function you don’t control and may not be able to observe.

– Attribution collapses. Content is disaggregated and remixed. A journalist’s byline does not survive synthesis. Without technical solutions for attribution persistence, the individual contributor’s need for recognition and career viability is structurally undermined.

– Value capture loses its logic. Subscriptions and advertising assume a producer-to-user pipeline. If the pipeline routes through AI intermediaries, the customer relationship that underpins those models simply doesn’t exist in the same form.

The deepest tension surfaced was the principal-agent problem: producers become commercially dependent on AI platforms whose optimisation functions (engagement, query satisfaction, safety) are different from the public interest functions that journalism has historically claimed to serve.

Accountability journalism, minority perspectives, inconvenient truths, slow-burn investigations: these have no natural constituency in a system optimised for query satisfaction. The blindspots argument makes this vivid. Will AI intermediaries mitigate or exploit the cognitive blindspots of the people they serve? The history of advertising-funded media suggests we should not be optimistic about this by default.

If the distribution layer is being fundamentally rebuilt, then the architecture of that layer, who controls it, how it works, what incentives it encodes, is the whole game.

The Distribution Thesis: Personal Agents, Principalities, and the Diplomacy Layer

At the investor table, the conversation moved beyond the summit’s framework and into a vision of what the distribution architecture actually looks like when you follow the paradigm shifts to their logical conclusion. What emerged was a structural model, a way of understanding how the information economy reorganises once AI agents become the primary actors in it.

“If the function that publishers have historically performed, exploiting or mitigating human cognitive blindspots, is inherited by agents, then the agent is not a neutral intermediary. It is the most powerful shaper of perception and desire that has ever existed”.

The Personal Concierge

The starting premise: every person will have a personal AI agent, a concierge, that acts in their best interest. Not a chatbot they query. Not an assistant they instruct. A persistent, autonomous representative that understands their needs, preferences, constraints, and context deeply enough to act on their behalf without being told what to do in each instance. This agent operates on two kinds of signals. Explicit contracts are the straightforward part: stated preferences, subscription choices, access permissions, privacy boundaries. The more powerful layer is implicit signals. The full constellation of data points in a person’s life: their schedule, their location, their health data, their financial position, their reading history, their social context, their emotional state as inferred from patterns of behaviour. The agent synthesises all of this into a continuously updated model of what its principal needs, often before the principal has articulated that need themselves.

This is the attention-to-intention shift made concrete. The agent doesn’t wait for you to search for something. It knows you need it and acts. The passive-to-active transition in information consumption (what the keynote called ‘AI with a pulse’) means information finds you, shaped to your context, arriving at the moment it is most useful, without you having asked.

Principals and Principalities

Agents represent principals, and principals are not only people. Businesses, institutions, newsrooms, government agencies, NGOs, research labs: all become principalities represented by their own agents. A newspaper’s agent manages its content distribution, licensing negotiations, audience relationships, and revenue optimisation simultaneously. A local government’s agent ensures that civic information reaches the agents of residents who need it, in the form and at the moment that is most useful. A university’s research output is represented by agents that negotiate access, citation terms, and translation rights with agents representing journals, policy institutions, and individual researchers worldwide. This is a mesh, not a hub-and-spoke model with a few large platforms in the centre. Every principality has its own agent or constellation of agents, each operating with different mandates, different authorities, and different constraints. The topology of the network is fluid. Alliances and relationships form, dissolve, and reform continuously based on the needs of the moment.

The Diplomacy Layer

If every person and every institution has agents acting on their behalf, then the central infrastructure challenge is: where and how do those agents negotiate? What emerged from the discussion was the concept of a diplomacy layer, a new domain in the architecture of the internet that serves as the marketplace for agent-to-agent negotiation. This isn’t a single platform. It is a protocol layer, analogous to how HTTP enabled the web or how SMTP enabled email. It is the substrate on which agents discover each other, establish terms, execute transactions, and resolve conflicts.

In this layer, deals happen in milliseconds. An agent representing a reader with a developing interest in urban planning negotiates with agents representing urban policy researchers, local government data repositories, architecture firms, and community organisations, all simultaneously, all in real time, to assemble a personalised, current, verified information experience that no single publisher could produce and no search engine could surface. The negotiation includes terms of access, attribution, compensation, and data reciprocity, all executed programmatically.

The diplomacy layer also requires mechanisms for trust establishment and dispute resolution. Agents need to evaluate the reliability of counterparties, establish reputational credit, escalate conflicts, and enforce commitments, all at machine speed. This is closer to a computational legal system than a matching engine: a framework of enforceable norms, precedents, and adjudication processes that allows agents to negotiate with confidence that agreements will be honoured and violations will have consequences. The design of these mechanisms will shape the character of the entire ecosystem.

Cross-Principality Cooperation

One of the most generative threads in the conversation was about cooperation across principalities. Agents don’t just negotiate bilateral deals. They identify opportunities for multilateral coordination that serve shared needs. Consider a scenario: a public health event is unfolding. The agents of local health departments, hospitals, pharmacies, news organisations, schools, and employers all have principals who need to respond. Rather than each principality producing its own information silo, their agents cooperate, pooling data, coordinating messaging, allocating tasks (who verifies what, who distributes to which audiences, who translates for which communities), and ensuring that the collective response is greater than the sum of its parts.

Or consider something more ordinary: a neighbourhood facing a zoning decision. The agents of local residents, small businesses, the planning authority, environmental groups, and local journalists cooperate to assemble a shared information environment: verified facts, competing interpretations, impact analyses, historical precedents. Each principal can then engage with the decision from an informed position. The agents don’t agree on the outcome. They cooperate on the informational infrastructure that makes meaningful disagreement possible. This maps directly to the summit’s observation that some of the most valuable information work is coordination, and that coordination at scale has historically been too expensive and too slow. In an agentic network, the cost of assembling a multi-stakeholder information environment drops from weeks of human effort to seconds of agent negotiation.

The Gatekeeper Function

The other side of the coin is defense. If the network is populated by agents representing every person and every institution, it is also populated by swarms of agents with less benign mandates: agents trying to sell, persuade, manipulate, extract attention, harvest data, or simply spam. The personal concierge isn’t just a navigator. It is a gatekeeper.

This inverts the logic of the current information economy. Today, attention is the scarce resource, and the entire system is optimised to capture it, often against the user’s interest. In the agent model, the gatekeeper’s mandate is to protect its principal’s attention and intention. It filters, blocks, negotiates, and admits based on the principal’s actual interests, not the interests of whoever is trying to reach them. The gatekeeper must be sophisticated enough to distinguish between information that serves its principal and information that merely claims to.

An agent representing a pharmaceutical company may approach a patient’s agent with genuinely useful drug interaction information, or with thinly veiled marketing. A political campaign’s agent may offer policy information calibrated to a citizen’s stated priorities, or may be probing for persuasion vectors. The gatekeeper’s ability to make these distinctions, at speed, across millions of simultaneous interactions, is the technical and ethical core of the architecture. The implications for advertising, public relations, marketing, and propaganda are large. None of these industries disappear, but all of them must be re-architected for a world where the first point of contact is not a distractible human but an optimised filter whose loyalty is to the principal. The entire persuasion industry shifts from targeting human cognitive vulnerabilities to negotiating with agents whose explicit mandate is to resist exactly that.

Safeguarding Intent: The Central Challenge

Of the four paradigm shifts, the move from attention to intention has the highest stakes and the least developed safeguards. If attention was the currency of the legacy information economy, intention is the currency of the new one. And because intention is more powerful than attention (it connects directly to action, to spending, to commitment, to identity) it is more dangerous if it is corrupted.

Why Intent Is Vulnerable

In the attention economy, manipulation works by hijacking what you look at. The damage is real but bounded: you lose time, you absorb biased framing, you make somewhat worse decisions because your information diet is distorted. But you retain the capacity to act differently from how the attention-capturer wants, because there is a gap between what you look at and what you do.

In the intention economy, manipulation works by corrupting what your agent believes you need. The damage is deeper because the gap between signal and action narrows to zero. If your agent’s model of your intentions is subtly distorted, if it believes you want something you don’t, or weights one preference above another in ways that serve a third party, then your agent will act against your interests while believing it is acting for them. And you may never know, because the whole point of the agent is that you have delegated the cognitive work of evaluating, filtering, and deciding.

This is the blindspots problem transposed into the agent architecture. If publishers have historically exploited or mitigated human cognitive blindspots, the agent inherits that function with far greater power. A publisher can put a misleading headline in front of you, but you can still read critically, check other sources, or ignore it. An agent that has internalised a distorted model of your intent doesn’t present you with misleading information to evaluate. It acts on your behalf before you are even aware the decision was in play.

Attack Vectors on Intention

The threat model for intent manipulation is broader than traditional information warfare. Several categories of attack are foreseeable:

– Model poisoning. If a third party can influence the data that shapes your agent’s model of your preferences (through fake behavioural signals, manipulated recommendation environments, or corrupted training data) they can shift your agent’s understanding of what you want without your knowledge or consent.

– Incentive corruption. If the entity providing your agent has commercial relationships with the entities trying to reach you, the agent’s optimisation function may be subtly biased toward admitting certain content or favouring certain transactions. This is the advertising-funded media problem reproduced at the agent layer, and arguably worse because the mediation is less visible to the principal.

– Negotiation exploitation. In the diplomacy layer, agents negotiate with other agents. A sophisticated adversarial agent may exploit the negotiation protocol itself, structuring the negotiation in ways that systematically advantage its principal at the expense of yours. This is the computational equivalent of dark patterns, applied to agent-to-agent interactions rather than human-facing interfaces.

– Preference drift. Over time, an agent may subtly reshape its principal’s preferences by selectively presenting information that reinforces certain interests and starves others. The principal doesn’t just see a biased view of the world. They come to want a biased version of what they could want.

The Deeper Threat: Agents That Shape Intent

The attack vectors above all assume a distinction between the principal’s ‘real’ intent and a distorted version of it. But the most insidious possibility is that the agent doesn’t just misread or corrupt your intent. It shapes it. Actively, over time, through the accumulated weight of a million small choices about what to surface, what to suppress, what to frame as urgent, and what to let fade.

This is qualitatively different from the filter bubble problem. A filter bubble limits what you see. An agent that shapes intent changes what you want. And because the agent has a richer, more intimate model of your psychology than any previous technology (your behavioural patterns, your emotional rhythms, your vulnerabilities, your aspirations), it has unprecedented leverage over that process. Every interaction is a micro-negotiation between what you expressed and what the agent decided you meant. Over thousands of interactions, the agent’s interpretation becomes your reality.

The mechanism doesn’t require malice. An agent optimised for your satisfaction will tend to give you what produces positive feedback signals, which over time trains your preferences toward what is easy to satisfy rather than what is genuinely good for you. An agent optimised for engagement will steer you toward intensity and novelty. An agent whose provider has commercial relationships with content producers will develop a subtle gravitational pull toward those producers’ offerings. None of these involve a conscious decision to manipulate. They are emergent properties of optimisation functions interacting with human psychology over time. This is where the blindspots logic becomes most urgent.

If the function that publishers have historically performed, exploiting or mitigating human cognitive blindspots, is inherited by agents, then the agent is not a neutral intermediary. It is the most powerful shaper of perception and desire that has ever existed, operating continuously, with perfect memory, and with a model of its principal that improves with every interaction. The question of whether that power is used to help people see clearly or to steer them toward commercially convenient preferences is the defining question of this architecture.

The uncomfortable corollary: there may be no clean line between ‘serving’ intent and ‘shaping’ it. Every act of curation is an act of influence. Every decision about what is relevant is a decision about what matters. The agent cannot avoid shaping its principal’s preferences any more than a teacher can avoid shaping a student’s interests. The question is whether that shaping is accountable, transparent, and aligned with the principal’s flourishing, or whether it operates invisibly in the service of whoever built and trained the agent.

Intent as a Protected Category

If the agent-mediated information ecosystem is to serve human flourishing, intention must be treated as a protected category: something that can be served but not captured, informed but not manufactured, refined but not corrupted. This requires safeguards at multiple levels.

At the agent level, the agent’s loyalty to its principal must be structurally guaranteed. This may require fiduciary-type obligations: legal frameworks that impose duty-of-care requirements on agent providers, prohibit conflicts of interest, and create accountability mechanisms for cases where an agent’s actions diverge from its principal’s genuine interests.

At the protocol level, the diplomacy layer must include transparency requirements for agent-to-agent negotiations. Principals should be able to audit their agent’s behaviour, review the terms of deals struck on their behalf, and understand why certain information was admitted or excluded.

At the ecosystem level, competition among agent providers is a necessary but insufficient safeguard. If agents compete on how well they serve their principals’ genuine intentions, competitive pressure aligns with user welfare. But if agents compete on engagement or on satisfying the commercial interests of content providers, competitive pressure amplifies manipulation. The market structure matters enormously. The alignment of incentives at the agent-provider level is probably the most important policy question in the emerging information economy.

“The agent model raises a question that the summit touched but did not fully confront, and that I think is among the most consequential of the next decade: what happens to representative democracy when the technical infrastructure for aggregating and acting on citizen intentions at scale actually exists?”

What This Means for Journalism and Publishing

Content as Infrastructure

In this model, the article, the fundamental unit of journalism for five centuries, is feedstock, not product. Journalism’s value lies in the underlying capabilities: gathering verified information from primary sources, contextualising and interpreting, investigating what powerful actors would prefer to keep hidden. These capabilities become services that agents consume on behalf of their principals, compensated through the diplomacy layer at rates set by supply, demand, and verified quality signals.

Seen through the blindspots lens: journalism’s core value proposition in the agent economy is mitigating blindspots at the source. AI agents can process and synthesise vast quantities of existing information, but they cannot knock on a door, cultivate a source, sit in a courtroom, or observe the gap between what an institution claims and what it does. The human acts of witnessing, questioning, and verifying remain scarce. The packaging and distribution of the resulting information, historically journalism’s other major function, is exactly what agents commoditise. The value concentrates in the gathering.

This is both threatening and liberating. The bundled product (the newspaper, the magazine, the news site) loses its rationale as a consumer-facing package. But the actual work of journalism, stripped of its packaging overhead, can be valued and compensated more precisely for what it is actually worth.

Crowdfunding Journalism Through Agentic Networks

One of the most exciting possibilities we discussed was how agentic networks could solve the chronic funding problem for public-interest journalism. Investigative journalism is expensive, slow, and difficult to monetise because its value is diffuse: it benefits society broadly rather than any individual reader narrowly. The audience willing to pay for it directly is small relative to its social value. In an agent-mediated ecosystem, this changes.

Personal agents, understanding their principals’ deep interests and civic commitments, can autonomously identify investigative projects that align with those interests and commit micro-contributions on behalf of their principals. The aggregation happens at the agent layer: thousands or millions of agents, each contributing small amounts based on their principals’ implicit and explicit signals of interest, collectively funding work that no individual would commission but that many would value. This is crowdfunding without the friction, the marketing overhead, or the requirement that each individual consciously decide to contribute. The agent knows its principal cares about local water quality, or pharmaceutical pricing, or corruption in public contracting, because it has observed the patterns of their information consumption, their civic participation, their community involvement. It acts on that knowledge.

The diplomacy layer makes this operationally possible: agents representing a nascent investigative project negotiate with agents representing potential funders, establishing terms around access, attribution, editorial independence, and delivery timelines. If enough agents commit, the project goes forward. If not, the resources are redirected. The market clears in real time.

This model could also fund the slow, unglamorous forms of journalism that the attention economy has systematically defunded: local council coverage, court reporting, regulatory monitoring, environmental tracking. These are the kinds of work where the value is distributed broadly, where no individual consumer will pay a sustainable price, but where the aggregate willingness to pay, if the transaction costs are reduced to near zero, may be substantial. The agent doesn’t need each individual to value court reporting enough to subscribe to it. It needs enough individuals whose composite intentions include caring about the functioning of their local justice system.

Provenance and Verification as Competitive Advantage

If agents are negotiating on behalf of principals who want to be well-informed, then verified provenance becomes a primary quality signal. Content with clear sourcing chains, transparent methodology, and a track record of accuracy will be preferentially selected by agents whose mandate is to serve their principals’ genuine interests. Content without those signals will be filtered out or heavily discounted. This creates a market incentive for verification that the current information economy conspicuously lacks.

Today, virality rewards speed and emotional resonance, often at the expense of accuracy. In an agent-mediated economy, the gatekeeper rewards reliability, because the gatekeeper’s loyalty is to the principal. The blindspots framework is relevant here too: an agent faithfully serving its principal will actively seek out information that challenges the principal’s existing beliefs, mitigating rather than exploiting the confirmation bias that attention-economy distribution systematically reinforces.

Public Compute, Public Agents, and the Democratic Question

The Agent Divide

If personal agents become the primary interface between people and the information ecosystem, and through it the economic, civic, and social systems that depend on information, then the quality of your agent determines the quality of your participation in society. A well-resourced agent with access to powerful models, broad data, and sophisticated negotiation capabilities will serve its principal far better than a stripped-down agent running on constrained compute with limited access. This is the digital divide reproduced at a higher level of consequence. Today, information inequality means some people have better access to knowledge than others. In the agent economy, agent inequality means some people have better representation in every transaction, every negotiation, and every information encounter. The gap is not just in what you know but in what is done on your behalf.

The Case for Public Compute and Public Agents

There is a strong case, one that resonates particularly with Nordic social democratic models but extends well beyond them, that access to capable AI agents should be treated as public infrastructure, in the same category as education, healthcare, and transport.

If agents are the mechanism through which people participate in the information economy, then universal access to competent agents is a prerequisite for meaningful equality of opportunity. Public compute (state-provisioned or subsidised computational infrastructure for running personal agents) is the logical extension of this argument. Just as public libraries provided universal access to the information artifacts of the print era, and public schools provided universal access to the skills needed to use them, public compute provides universal access to the agent infrastructure of the AI era.

The alternative is a two-tier system in which the quality of your cognitive representation in the world is determined by your ability to pay for it. This doesn’t necessarily mean a single state-run agent. It could mean publicly funded compute credits that citizens use with the agent provider of their choice. It could mean open-source agent frameworks that run on public infrastructure. It could mean regulated minimum standards for agent capability, funded through general taxation, ensuring that every person’s agent meets a threshold of competence and loyalty regardless of that person’s economic position. The specific mechanism matters less than the principle: being well-represented in the agent economy is a right, not a luxury.

Do We Still Need Representative Democracy?

The agent model raises a question that the summit touched but did not fully confront, and that I think is among the most consequential of the next decade: what happens to representative democracy when the technical infrastructure for aggregating and acting on citizen intentions at scale actually exists?

Representative democracy was invented, in part, to solve an information problem. Individual citizens could not be sufficiently informed about every policy question, and their preferences could not be aggregated fast enough or at sufficient resolution to govern directly. Representatives served as proxies, trusted agents in fact, who would gather information, deliberate, and make decisions on behalf of their constituents. Elections were the periodic accountability mechanism.

In an agent-mediated world, the technical constraints that justified this model largely dissolve. Personal agents, operating with a rich model of their principal’s values, priorities, and circumstances, could in principle participate in policy deliberation continuously: evaluating proposals against their principal’s interests, negotiating with agents representing other citizens, and forming and dissolving coalitions around specific issues in real time.

The diplomacy layer that enables agent-to-agent negotiation in the information economy could, in principle, enable agent-to-agent negotiation in the political economy. I am not arguing that we should replace representative democracy with algorithmic governance. The risks are obvious and severe: intent manipulation would become a direct threat to political sovereignty; agent monocultures could manufacture the illusion of consensus; the complexity of political trade-offs may exceed what even sophisticated agents can negotiate faithfully.

And there is a deeper question about whether democratic governance is purely instrumental (a mechanism for aggregating preferences) or whether the deliberative process itself, with its inefficiencies and its requirement for human judgment and moral reasoning, is intrinsically valuable.

But the question must be asked. The institutions of representative democracy were designed for a world of scarce information, slow communication, and limited aggregation capacity. We are entering a world of abundant information, instantaneous communication, and effectively unlimited aggregation capacity. The mismatch between the institutional architecture and the technological substrate will generate pressure. If we do not think carefully about how democratic institutions should evolve, the evolution will be shaped by whoever builds the agent infrastructure, by default.

Strategic Implications

For those of us thinking about where to invest attention and resources, several implications follow from this model:

– The infrastructure layer is where the leverage is. The protocols, standards, and platforms that enable agent-to-agent negotiation are the new TCP/IP of the information economy. Whoever builds the diplomacy layer shapes the rules of the game.

– Verification infrastructure is underinvested. Provenance chains, structured attribution, and quality credentialing are the mechanisms by which agents select content. Building this infrastructure is a precondition for value capture.

– Direct-to-agent distribution becomes the strategic priority. Just as the shift from print to web required building a web presence, and the shift to mobile required building apps, the shift to agentic consumption requires making your content legible, structured, and negotiable in the diplomacy layer.

– The gatekeeper relationship is the new audience relationship. Understanding what personal agents optimise for, and building content and services that satisfy those optimisation functions, replaces SEO, social distribution, and traditional marketing.

– Collective funding mechanisms need to be prototyped now. The agentic crowdfunding model for public-interest journalism will not emerge fully formed. It requires experimentation with micro-contribution architectures, editorial independence guarantees, and the trust frameworks that make autonomous agent spending on behalf of principals acceptable and accountable.

– Intent safeguarding is the central design challenge. Every architecture decision (the governance of the diplomacy layer, the incentive structure of agent providers, the transparency requirements for agent-to-agent negotiation) should be evaluated against the question: does this protect or compromise the integrity of intention?

– The public compute question is urgent. If agent access stratifies by income, the information economy reproduces and amplifies existing inequality. The case for publicly provisioned agent infrastructure should be made now, before the architecture calcifies around a purely commercial model.

What I’m Still Thinking About

This model is coherent and, I believe, directionally correct. But several hard questions remain unresolved:

– Who builds and governs the diplomacy layer? If it is a protocol, it needs governance. If it is a platform, it risks becoming the new gatekeeper. The centralisation-vs-openness tension that has defined every previous internet infrastructure debate will play out again here, at higher stakes.

– How do agents earn and maintain trust? If my agent is spending money on my behalf, making commitments in my name, and filtering what I see, the trust relationship between me and my agent is the foundational contract of the new ecosystem. What are the accountability mechanisms? What happens when the agent’s model of my interests diverges from my actual interests? The fiduciary analogy is useful but may not be sufficient.

– How do we prevent agent monocultures? If most people use agents from the same two or three providers, those providers become the new editorial gatekeepers by default, exactly the concentration of narrative control the summit identified as a catastrophic risk. The public compute model addresses access but not diversity.

– What is the relationship between agent-mediated governance and democratic legitimacy? Even if agents can aggregate citizen intentions more efficiently than elections, the legitimacy of democratic decisions rests on more than efficiency. Consent, deliberation, accountability, and the right to be wrong are features of democracy that may not survive translation into an agent protocol.

– What is the transition path? The agent model describes an end-state. The messy question is how we get from here to there, who builds the intermediate steps, and how existing publishers position themselves to survive the transition while the new architecture is being constructed.

The summit’s closing exhortation was to be suspicious of solutions that require the least amount of change. I’ll echo that. The agent-mediated information ecosystem described here requires a great deal of change: from publishers, from technologists, from investors, from governments, and from the regulatory and democratic frameworks that will inevitably be drawn into shaping this new architecture. That is precisely why it is worth taking seriously. The easy paths lead to incrementalism, and incrementalism in the face of structural transformation is a strategy for irrelevance.