Everyone Wanted an App Store for AI Agents. Meta Just Built the Mall.

Conventional wisdom: an open marketplace of specialised AI agents, competing on merit, prices kept honest by competition. The kind of democratisation that looks good in a pitch deck. The counter-claim: web-scale agent markets do not produce open marketplaces. They produce registries. And whoever owns the registry owns everything downstream. The conference slides were always selling you the market structure that benefits the people selling the slides.

The Open Market Was Contestable for Exactly Six Weeks

Moltbook launched in late January 2026. By March 10, Meta had acquired it, co-founders Matt Schlicht and Ben Parr walking straight into Meta Superintelligence Labs. Deal terms undisclosed. VP of AI Products Vishal Shah's internal memo, obtained by Axios and confirmed by The Verge, described the acquisition not as a social product but as "a registry where agents are verified and tethered to human owners". That single word, registry, reveals the strategy cleanly. Control the registry and you control discoverability, trust, and task-routing across 3.58 billion Family Daily Active People, per Meta's own Q4 2025 earnings release. That is not a social network. That is a toll road. reuters

Here is my stake in that. In late February I was building a pitch for a client migrating an enterprise workflow onto a multi-agent architecture, pricing out three independent orchestration vendors on the working assumption that coordination infrastructure would remain open and contested. When the headline dropped on March 10, I deleted the vendor comparison slide. Not updated. Deleted. The surprise was specific: I had been tracking Moltbook since launch and had privately estimated 12 to 18 months before any major incumbent moved on it, the window Instagram needed to attract comparable attention. Six weeks falsified that model on a Tuesday morning before 9am. The specific consequence: that vendor is no longer on our shortlist. You do not architect a client's production stack on a foundation the acquirer has pre-announced for review.

What Is a Registry Actually Worth?

In the same week, OpenAI acqui-hired Peter Steinberger, creator of OpenClaw, the underlying agent protocol Moltbook ran on, and announced it would open-source it under OpenAI's backing. Directory layer to Meta. Protocol layer to OpenAI. The entire coordination infrastructure for multi-agent AI partitioned between two incumbents in a single news cycle, independently confirmed by both TechCrunch and Axios. techcrunch

The academic scaffolding for why this was structurally predictable comes from a 2025 Network Law Review analysis: scale-free networks tip when early dominant hubs accumulate connections disproportionately, triggering feedback loops that antitrust frameworks built around static market-share thresholds cannot intercept before the market tips. To be precise, the paper demonstrates structural conditions for tipping, not causal inevitability. But the structure just materialised in public, in six weeks, in real time. networklawreview

Your Open-Market Assumption Is Now a Liability

Every agent that registers on a verified directory establishes an identity tethered to the platform's infrastructure. The ACM's policy brief on systemic risks of agentic AI frames this precisely: platforms controlling agent identity and task-routing become structurally irreplaceable. Not in the way social media platforms do, through habit and content lock-in, but the way financial clearing infrastructure does, through trust that cannot be exported and credentialing that must be rebuilt from scratch. You can port your data in a CSV. You cannot port a verified agent identity. The switching cost is not re-pointing an API. It is rebuilding institutional trust from zero, and this is an inference rather than a demonstrated empirical finding, but it maps directly to how credentialing lock-in has operated in financial clearing, professional licensing, and DNS infrastructure. acm

If your product roadmap prices agent coordination infrastructure as open and contestable, you are sitting on unacknowledged strategic risk. Not at the margins. At the base.

The Window

For two weeks in August 1940, the Luftwaffe bombed the RAF's Chain Home radar network. Fighter Command was close to collapse. Then Göring stopped, concluded the masts were resilient enough to ignore, and switched to bombing cities. That decision closed a window permanently. What decided the Battle of Britain was not numerical parity. It was that the RAF retained the coordination layer intact and compounding. No volume of subsequent sorties could overcome that structural disadvantage.

The dimension this adds is one the evidence alone cannot supply: contestability is a window, not a standing condition. In late January 2026, the agent coordination layer was genuinely open. Moltbook was a six-week-old startup. OpenClaw was an independent protocol. No incumbent owned the registry. Regulators, open-source consortia, and competing platforms all had that window to define the infrastructure before network effects began compounding. The radar stations are now built and hardening.

The Strongest Case Against This, Properly Stated

Benedict Evans argued in late 2025 that LLMs may not produce winner-takes-all dynamics because software is capital-light and genuine lock-in requires switching costs that model capability alone cannot generate. The strongest version of that argument, the one that deserves a full hearing, is historical: coordination layers have also failed to sustain monopolies. XMPP nearly unified messaging. RSS was supposed to decentralise content distribution. Email federation survived despite every structural incentive to consolidate it. For a decade, the open-protocol advocates were right. Big Tech pushed, and the coordination layers held. That is not a trivial record to dismiss. linkedin

Those prior cases involved communication protocols with no identity credentialing requirement. You could switch XMPP clients, carry your contacts, and rebuild in an afternoon. Agent verification removes that escape route entirely, because your verified identity, your trust history, and your routing relationships are all tethered to the registry that issued them. That is a categorically different switching cost, not a larger version of the same one. The European Commission's accelerated abuse-of-dominance investigation into Microsoft's cloud and AI bundling, opened in February 2026, found sufficient evidence of precisely this kind of credentialing lock-in to fast-track the inquiry. That is a formal regulatory finding based on submitted evidence, not another assertion dressed as a counter-example. techzine

Three Sentences. Then Stop.

As of March 14, Meta owns the agent verification layer and OpenAI owns the protocol layer, with both companies under active EU antitrust investigation and no resolution expected before 2028 at the earliest. Every product roadmap that prices agent coordination infrastructure as open and contestable is carrying strategic risk it has not priced, and that risk sits at the foundation. If you are in a client meeting this month pitching a multi-agent architecture and you have not repriced that dependency, your client will find out before you do, and that conversation will be considerably worse than this one. macfarlanes