I’m super-empowered by my Claw and excited for the future. But I want to be realistic about what will stop the agentic web from breaking bad. So I built a simulation model to answer that question…

By Tony Curzon Price

TL;DR

Platforms win not by being evil, but by being genuinely useful first. Google, Facebook, and others captured open ecosystems (email, the web) by extracting ecosystem-level social signals — like spam patterns or social graphs — and feeding them back as better services. No individual user could do that alone. That real value is step zero of enshittification.

The agentic web faces the exact same trap. Right now it’s beautifully open and decentralised, but it has no coordinated way to handle security or discovery across the ecosystem. Apple, Google, or Meta will offer managed agent platforms that solve those problems — genuinely better at first — and that’s how capture begins.

The fix: build a “Social Signal Processing Layer” — not a platform, but commons-based institutions like FIDU (a data union with absolute loyalty to members) that can aggregate trust and discovery signals without the structural incentive to enshittify. The simulation model Enshait shows the single most important variable is how fast these institutions mature. Act too slowly and platforms lock in; act early and the open ecosystem gets a real shot.

We also need portability regulation (so switching stays real), antitrust pressure (to slow platform entry and buy time), and agent capabilities that reduce the friction of decentralised protocols. No single lever is enough — the decentralisation window is real and finite.

Here goes for the full text. (I used the FIDU Chatlab to help me write this – I describe in detail how at the bottom of the post.)

Background and thanks

I’ve just spent a wonderful week in the US – first thanks to Shuwei Fang at her excellent Signals at Scale workshop at the Kennedy School, which was inspiring for the optimism and determination in the room to make the AI revolution good for media ecosystems. Then I spent the weekend with Gigi Danziger and Albert Wenger at their farm near Hudson. Gigi and Albert are behind the amazing Eutopia Foundation that has renewed its funding for The First International Data Union (FIDU). Thank you!

But more than that, conversations during long walks in upstate NY and the drive back to NYC (most of it on Tesla self-drive… first time for me & I was very impressed) were deep and probing about the extraordinary moment we are living through as AI and agents rewire the web (and eventually rewire society … again).

This post – and enshAIt, the discrete time dynamic logistic modal choice model that I have developed and released to explore scenarios for the evolution of the agentosphere – are the product of those conversations.

=====================================================================

Remember email before Google?

Before Gmail, before Hotmail, before any of them — email was just a protocol. SMTP, IMAP, POP3. You ran your own server, or your university ran one, or your ISP gave you a mailbox. Nobody owned email. Nobody could. It was a set of open standards, and anyone who implemented them could talk to anyone else who implemented them. It was glorious.

And then it wasn’t.

Not because the protocols failed. They’re still there. You can still run your own mail server today. About two per cent of people do. The rest use Gmail, Outlook, or one of a handful of providers. The protocols won the architecture and lost the users.

Why?

The General Theory of Platform Market Power

Here is a claim I want you to take seriously, because I think it explains nearly everything about how open ecosystems become captured ones: platform market power always rests on the extraction and redistribution of social signals.

Not data. Not compute. Not even network effects in the abstract. The specific mechanism is this: a platform observes the behaviour of all the nodes in a network, extracts a signal that no individual node could compute on its own, and feeds that signal back into the network as improved service. That improvement is real. It is genuinely valuable. And it is the first step toward capture.

PageRank is the clearest example. The early web was a mess of pages with no way to assess quality. Google looked at the link structure — a social signal, the collective editorial judgement of millions of webmasters — and fed it back as relevance-ranked search results. This was not a trick. It was a genuine public good, for a while. Every individual web user benefited from Google’s ability to read the whole graph. No individual user could have done that for themselves.

Facebook’s News Feed did the same thing with the social graph. Early Facebook looked at your connections, your interactions, your friends’ interactions, and surfaced content you were likely to care about. Before the feed, you had to manually visit each friend’s profile page. The feed was better. It was better because Facebook could see the whole network and you couldn’t.

This is, I believe, a general theory. Platform market power does not begin with exploitation. It begins with a genuine service improvement that only an entity with a view of the whole ecosystem can provide. The platform makes the ecosystem better. Then it captures the ecosystem. Then it makes the ecosystem worse. Cory Doctorow’s enshittification thesis describes the second and third steps brilliantly. But the first step — the one where the platform is actually right that it’s providing value — is the one we need to understand if we want to prevent the cycle from starting.

Because you cannot fight step one by saying the platform is wrong. At step one, the platform is right.

The Agentosphere Is Wide Open — and Wide Open Is Not Enough

Now look at what is happening in the agentic web.

We are in the middle of what Jonathan Zittrain might call a generative explosion — or what a palaeontologist would recognise as a Cambrian radiation. The agentosphere is erupting. OpenClaw and similar open frameworks have made it trivially easy to build, deploy, and connect agents. A2A protocols are emerging. MCP is giving agents a shared way to interact with tools. The energy is extraordinary. Thousands of developers are building agents that can negotiate, transact, discover, and collaborate with other agents across an open ecosystem.

It feels like the early web. It feels like email in 1995. The sun is shining, everything is open, creativity is everywhere, and the future belongs to protocols, not platforms.

I have been here before. So have you, if you are old enough. The dangers should be clear.

What Decentralisation Does Badly

Pure decentralisation — protocols without institutions — has a specific and well-documented failure mode. It cannot efficiently process ecosystem-level signals and feed them back to individual nodes.

Consider what happened to email.

Those of us who ran our own mail servers in the 2000s remember the experience vividly. We were drowning in spam. Our outbound mail was being blocked by recipients whose providers didn’t recognise our servers. We spent hours configuring SPF records, DKIM signatures, greylisting, and Bayesian filters. It was a losing battle, because spam detection is fundamentally an ecosystem-level problem. You need to see the patterns across millions of mailboxes to distinguish spam from legitimate mail. No individual server operator could do that.

Gmail solved it. Not through any proprietary protocol — Gmail speaks SMTP like everyone else — but through social signal processing at scale. Google could see the spam patterns across hundreds of millions of inboxes and feed that intelligence back to each individual user. The result was transformative. Gmail’s spam filtering was so much better than anything a self-hosted server could achieve that it became, for most people, the rational choice. Not because open email was bad in principle, but because open email couldn’t solve its own signal-processing problem.

Discovery was the other failure. Email couldn’t tell you who you should be talking to. It couldn’t surface relevant connections or prioritise your inbox based on what actually mattered to you. That vacuum created the opening for Facebook, LinkedIn, and every social platform that followed. They could look at the whole social graph. You couldn’t.

Protocol-level decentralisation, without a social signal-processing layer, loses to platforms. Every time. Not because protocols are inferior, but because they are incomplete.

The Agentosphere Will Face Exactly the Same Pressures

Now transpose this to the agentic web, and the parallels are almost painfully exact.

Security. Right now, the open agentosphere has no coordinated mechanism for identifying malicious agents, compromised tools, or fraudulent A2A interactions. Each agent — or each agent’s human — is on their own. This is fine while the ecosystem is small and populated mainly by technically sophisticated early adopters who can evaluate risks themselves. It will not be fine when hundreds of millions of non-technical users have agents acting on their behalf, negotiating contracts, sharing data, and making purchases.

Imagine Apple’s agent offering. Apple already has a billion devices, a security infrastructure, and a brand built on the promise of protection. An Apple agent platform would have oversight of the security landscape across all its users’ agents. It would detect fraud patterns, block malicious agents proactively, and update protections in real time. It would be genuinely better at security than any individual open agent could be on its own. For everyone except the most technically capable users, this would be a real and rational improvement.

Discovery. The open agentosphere currently has no good mechanism for one agent to find another agent that its human might benefit from knowing about. A2A protocols let agents talk to each other, but they don’t help agents find each other. There is no equivalent of search, no equivalent of a social feed, no way for the ecosystem to surface relevant connections.

Imagine Meta’s agent offering. Meta has the social graph. A Meta agent platform could parse A2A interactions across its entire user base and say: “This agent has information that that agent’s human would probably find valuable. Let us facilitate the introduction and negotiate the information exchange.” Before enshittification — and this is the crucial qualifier — that would be a genuine improvement on the open alternative, where agents have no way to discover each other beyond whatever their humans manually configure.

These are not hypothetical threats. They are the obvious, rational, and in some sense correct moves for platform companies to make. The first company to offer a managed agent platform with ecosystem-level security and discovery will provide a genuinely better experience for most users. That is step one of the enshittification cycle. And if we have learned anything from the last thirty years, it is that step one is where the battle is won or lost.

So What Stops This?

If the diagnosis is correct — that platformisation happens because open ecosystems cannot process their own social signals — then the prescription follows directly.

The agentosphere needs a layer above the open protocol layer: a Social Signal Processing Layer.

This is not a platform. It is not a company. It is not a new protocol. It is an institution — or a group of institutions — that can do what platforms do (aggregate ecosystem-level signals and feed them back to individual nodes) without doing what platforms inevitably do next (capture those signals as a source of market power and extract rent from them).

What would such an institution look like? It would need to:

• Steward members’ data with an absolute duty of loyalty to them — not to advertisers, not to shareholders, not to its own growth metrics
• Aggregate trust and security signals across the open ecosystem — seeing enough of the network to detect threats, without becoming a surveillance chokepoint
• Enable discovery — helping agents find relevant counterparts, without controlling or biasing the matching
• Feed signals back as a commons — the processed intelligence goes to members’ agents, not into a proprietary moat

This is what the First International Data Union (FIDU) is striving to be. FIDU is a not-for-profit, purpose-driven organisation that stewards members’ data with an absolute duty of loyalty to them. It ensures data is used in accordance with members’ values. In the context of the agentosphere, FIDU represents the kind of institution that could provide the social signal processing layer — security intelligence, discovery facilitation, trust attestation — without the structural incentive to enshittify.

The critical question is whether institutions like FIDU can mature fast enough. Because the platforms are not waiting.

Modelling the Race: Introducing Enshait

To explore this question rigorously, I built a simulation model called Enshait — a portmanteau that should need no explanation.

Enshait is a discrete-time stock-flow model that simulates how the agentic web splits between open, commons-aligned signal processing and centralised platform signal processing over an eight-year horizon. It is not a forecast. It is a structured argument — a way of making explicit the assumptions behind different stories about the future and seeing where those assumptions lead.

The model works like this. A population of potential adopters enters the agentic ecosystem over time, following staggered adoption curves. Each adopter chooses between the open architecture and the platform architecture based on a logistic choice function — essentially, they go where the utility is higher, with some noise. The utility of each option depends on several evolving state variables:

• Commons signal quality — how good is the open ecosystem’s trust, security, and discovery infrastructure? This depends on two things: the institutional maturity of commons-based signal processors (the FIDU variable, F(t)) and the scale of the open ecosystem. You need both institutions and users. Neither alone is sufficient.
• Platform signal quality — how good is the platform’s proprietary signal processing? This starts high (platforms have existing infrastructure and data), grows with the platform’s installed base, but eventually degrades as enshittification kicks in.
• Enshittification — modelled as a structural tendency that activates once the platform achieves a threshold share among adopters, ramps up over time, and is braked by competitive pressure from a viable open ecosystem. This captures the crucial feedback loop: platform dominance accelerates extraction, but a healthy open alternative restrains it.
• Agent friction reduction — the “agents aren’t humans” factor. AI agents can handle the complexity of decentralised protocols (cryptographic attestations, distributed reputation queries, federated discovery) far better than human users ever could. This is what makes the agentic web potentially different from the human web. On the human web, platforms won partly because they reduced friction. If agents can handle the friction of decentralisation, the platform’s convenience advantage shrinks.
• Lock-in — switching costs that accumulate once platforms achieve dominance, making it progressively harder for adopters to leave.
• Values alignment — the utility premium of knowing your data steward has a duty of loyalty to you, not to its shareholders. This starts small and grows with awareness.

Adopters are not homogeneous. The model includes four consumer types — Sovereigntists (high values sensitivity, early adopters), Pragmatic Techies (balanced evaluators, the swing vote), Convenience Seekers (go where the experience is best right now, the majority), and Reluctant Adopters (follow the largest installed base, arrive late). Each type has different sensitivity to signal quality, network effects, lock-in, enshittification, agent friction reduction, and values alignment. Their staggered arrival creates the temporal drama of the model: the early ecosystem is dominated by types who favour openness, but the mainstream wave brings types who favour convenience.

The Scenarios

The model includes five preset scenarios that tell different stories about the same structural dynamics. They differ in four key parameters: how fast commons institutions develop, how strong the platform’s head start is, how much agents reduce friction for the open side, and when enshittification kicks in.

Pre-agent web (email-like) is the baseline parable. Protocols are open and the early phases of the ecosystem are entirely open and decentralised. Social institutions develop slowly. Therefore the platform has a strong head start in social signal processing. The result feels like what actually happened to email: platforms own the default, the open ecosystem survives at the margins, perhaps ten per cent of adopters on open-leaning paths. This is the “do nothing different” from last time scenario.

Platform Capture is the stark version of where we’re headed. Slow institutions, early enshittification, weak agent relief. The open ecosystem is viable in principle but loses on the clock. Incumbency, timing, and rent extraction reinforce each other. By the time enshittification creates an opening, lock-in has closed it.

The Decentralisation Scenario is the optimistic case. Institutions develop fast — FIDU-like organisations reach maturity before the mainstream adoption wave. The platform’s head start is modest. Agents substantially flatten protocol friction. The open ecosystem gets a real shot before dynamics lock in. This is not “platforms are harmless” — it is “the open side is ready when it matters.”

Federated Equilibrium is the middle path. Institutions develop at a credible pace. Platforms are competitive but not overwhelming. Enshittification is delayed by competitive pressure. The result is coexistence and contestability — not automatic decentralisation, but rules and norms that keep extraction in check. Think of it as the email world if email had developed better spam filtering and discovery before Gmail arrived.

Late Reversal is the painful path. Institutions are slow, platforms dominate early, but agents are strong and enshittification eventually becomes severe enough to push adopters back toward the open ecosystem — if commons institutions have been quietly maturing in the margins and are ready to receive defectors. This is the Twitter-to-Bluesky story, but for the whole agentic web.

What the Model Shows

The single most important variable in the model is the speed of institutional development — how quickly commons-based signal processors like FIDU reach maturity. Under almost every combination of other parameters, this is what determines whether the outcome looks like Platform Capture or The Decentralisation Scenario.

This is not a coincidence. It follows directly from the diagnosis. If platform power comes from social signal processing, then the open ecosystem’s fate depends on whether it can provide its own social signal processing. And that requires institutions, not just protocols.

The model also shows something subtler and more important: the open ecosystem restrains platform behaviour even for platform users. When the open side is viable, the competitive brake on enshittification is engaged. Platforms degrade service more slowly because users have somewhere else to go. This means that investment in the commons benefits everyone, not just the people who use them. FIDU and institutions like it doesn’t just serve its members — it disciplines the entire ecosystem.

You can explore the model yourself over here. Move the sliders. Watch what happens when you speed up institutional development. Watch what happens when you slow it down. The window of opportunity for decentralisation is real, and it is finite.

The Levers We Need to Pull

The model makes the argument concrete, but the argument is simple enough to state without it:

The agentosphere will enshittify — will ensh-AI-ttify — unless we act on multiple fronts simultaneously, and act soon.

No single intervention is sufficient. The Decentralisation Scenario is a window precisely because several things need to come together before it closes. Here is what we need:

1. Commons-based signal processing institutions like FIDU — data unions and purpose-driven stewards that can aggregate trust, security, and discovery signals across the open ecosystem with an absolute duty of loyalty to their members. Not protocols alone. Not companies. Institutions with governance structures that make enshittification structurally impossible. These are the core of the social signal processing layer — the thing the open ecosystem is missing and that platforms will provide if we don’t. The time to build them is now, before the mainstream adoption wave, before it is “rational” in a narrow market sense. The model shows that institutional maturity at year three matters far more than institutional maturity at year seven.
2. Regulation mandating open access and data portability — if your agent’s data, relationships, and interaction history are locked inside a platform, switching is theoretical rather than real. Portability regulation — in the spirit of the EU’s Data Act and GDPR’s right to data portability, but extended explicitly to agentic data and A2A interaction histories — turns lock-in from a one-way ratchet into a manageable friction. In the model’s terms, this caps L(t): switching costs cannot accumulate without limit if regulation guarantees you can take your data and leave. Open access mandates — requiring platform agents to interoperate with open-protocol agents on non-discriminatory terms — directly boost the network effects available to the open ecosystem. If platform agents must speak open protocols, then N_open’s network value rises even for users who haven’t left the platform yet. This is the DMA logic applied to the agentosphere, and it needs to happen before platform architectures calcify.
3. Antitrust enforcement that slows platform entry — crucial to the model’s dynamic is the time that the open ecosystem has to develop before being platform-captured. In the model this can be directly simulated with the “Open adoption lead — platform entry delay (months)” slider. There are parallels here to how antitrust has worked in the previous rounds of technological transformation. IBM was famously nervous to enter the PC market with too heavy-a hand because regulators were after them for dominance in mainframes. That allowed Microsoft, Intel and a few others the breathing space to create a relatively more open and contestable space. Same with Microsoft and the 1990s Browser Wars – MS was being run by lawyers and tired of antitrust, and that moderated its legendary aggressiveness in the emerging web space. I have written about this effect here and my friend and ex-colleague Alex Chisholm put it all very well when CEO of the UK’s antitrust agency in a speech in 2014 over here. So the competition authorities should keep their general pressure on to give maximum power to lawyers in the board rooms of the platforms.

The Cambrian explosion of the agentosphere is beautiful. I do not want to dampen anyone’s enthusiasm. But the Cambrian explosion was followed by consolidation, extinction, and the emergence of a few dominant forms. That is the natural trajectory of generative ecosystems — unless we build the institutions that keep them open.

The sun is shining. The time to build the roof is now.

Enshait is open source and available here. Play with the model. Challenge the assumptions. Move the institutional development slider and watch the future change. Then join FIDU.

======================================================================

[Yes, this note was written using the FIDU ChatLab. I gave the ChatLab context on FIDU and the Enshait model then gave ChatLab the following instruction:

Can you produce a first draft of a blog post entitled “When will the agentosphere enshaittify and how to stop it”

INSTRUCTIONS/STRUCTURE

This should start with “Remember email before Google, Hotmail , etc…”

Then say: the key point about platform market power is that it rests in all cases on extracting a social signal and using – at first – to improve the social environment it’s working in. Give brief examples of why this applied to PageRank, to Facebook’s social graph. Propose it as a general theory of platforms market power.

Make the point that this is something crucial about step 1 in Doctorow’s Enshittification Journey – the platform makes something better before it captures it and makes it worse.

Then say the same will happen in the Agentosphere. [OK, we’ll be a bit more optimistic at the end, but it’s worth grappling with the extent and urgency of the danger].

Talk about the dramatic growth of the agentosphere, and OpenClaw in particular. All of us in this space are in a full-on “generative explosion” to combine terms from Zittrain and the paleontological record. The excitement is palpable. We are open, we are decentralised, we are being creative, the sun is shining.

BUT THE DANGERS SHOULD BE CLEAR.

There are things that pure decentralisation does badly – look at security and discovery. Relate that to the early days of email. Those of us running our own servers were overwhelmed with spam, and overwhelmed with others thinking we were spam and blocking us. That was like a security threat. Gmail solved it with social intelligence. Also, email was no good at discovery, so it made paved the way for Facebook and others who could look at the whole social graph and prioritise the inbox.

So protocol-only level decentralisation loses to social signal extracting platforms.

The same will be true of the agentosphere, and security and discovery are very good places to start. Imagine Apple’s agent offering. It will have oversight of the security traps in all the agents of all its users, and will be able to pro-actively block bad behaviour before it spreads. A real improvement for all but the techiest. Imagine Meta’s agent offering – it will create an A2A message parsing system saying “this agent has something that agent will find probably find interesting, and we can facilitate the negotiation between the agents to swap that informatoin”. Again – until that become manipulative, ie pre-enshittification – that will be an improvement on full decentralisation.

So what is needed to avoid platformisation and what that brings?

Well, the biggest thing we need is a layer above the open protocol layer that is the “Social Signal Processing Layer”. That is, we need to have organisations that can steward our own and our agents’ data according to our values and in our interests. That stewardship requires trustworhty data access, signal aggregation and signal passing to verified participating agents. Without this thin layer, platformisation will return. that is what FIDU is striving to be.

This is the point in the blog post where I introduce the Enshait model and describe the scenarios and parameters, and all the levers that we need to work with. Present the model, show a few key graphs from the scenarios and encourage people to play with it.