State of the Agent Economy — Q3 2026

A dated account of the agent economy — not a snapshot. Current-state views cannot see what is dying, what is migrating between standards, and what identifiers are being reused after the thing behind them died. This report restates facts from a continuous, entity-resolved census of 5.2M AI assets and pins each dated observation to a public hash chain (/verify). It ranks nothing, scores nothing, predicts nothing — it counts and dates. ← news summary

Method

A transition is a component whose observed state changed between two dated reads; a change inside an unobserved span is interval-censored — dated to the span, not a single day. States: alive (≤180d) / dormant (180–365d) / abandoned (>365d) / archived / deleted (HTTP 404) / not_observed (403/429/timeout, not death). A component we have not observed is not-checked, never gone. A 404 is deleted or made private, so we say “deleted/inaccessible.” Every dated observation is a line in a daily SHA-256 slice chaining to a public head. CC-BY-4.0.

1. MCP server mortality — death and abandonment are different numbers

Through the first half of 2026 a “~52% abandoned” line circulated, off samples of one to two thousand. We censused the whole public github-linked set — 28,824 repositories — and separated staleness from death. The live 5-state panel classifies the cohort as 12% deleted, 21% fully dead (deleted or abandoned), and 31% dead or no longer maintained (adding dormant + archived). The frozen v1 census (2026-07-22) independently lands at 29% dead-or-stale — two reads agree. “Abandoned” headlines measure staleness; death is a different, smaller number, and dead-or-anything tops out near a third, not a majority. Median lifespan is not reported — MCP births are left-censored, so we defer it rather than fabricate it.

2. Migration Watch — a dated baseline of a standard in motion

MCP's protocol is migrating (the 2026-07-28 revision changed discovery). Adoption is a wave measurable only from the moment capture starts. Of servers probed for protocol version, only 2.2% have adopted the current spec; 98% still speak an older generation, the long-dominant 2025-06-18 version holding ~81%. We have two observation dates so far and the probed population grew between them, so we withhold a cross-date migration rate and publish a dated baseline — the first fixed point from which real migration will be measured, not a rate we claim today.

3. Schema mutation & resurrection — what no snapshot can produce

Across the window the census recorded 998 dated MCP state transitions and 33,481 npm transitions. Beyond lifecycle, the tools themselves change: across 16 dated schema slices, 1,027 MCP servers changed a tool's declared schema — its input parameters or its description — while keeping the same tool name. An agent that resolved a tool's contract on one date could meet a different contract later, with no version bump to signal it; upgrade and rug-pull look identical until a human reads the diff, which is exactly why the dated record matters. We date the change and never infer intent.

Resurrection is rare. 3,510 distinct MCP servers have been observed deleted; only 6 ever returned to life — about 1 in 585. A dead identifier is almost never reclaimed by its original project, and when a name comes back it is often a different project reusing a vacated slot. For a consumer who pinned a dependency to that name, the identifier survived while the thing behind it changed — a discontinuity a current-state lookup hides entirely.

Ongoing incident evidence for named entities — dated register observations when an AI-agent security incident occurs — is at /incidents.

4. The crawl asymmetry — machines read the pages; the machine path sits unused

We instrument our own public census surface. Over a representative week it was fetched by roughly 9,500 machine requests a day against low-hundreds of human sessions, and answered by only ~22 structured API calls a day: machines hit the HTML pages about 438x more often than they use the structured API built for them — scraping HTML they must parse rather than reading the clean, dated, point-in-time API one link away. The consumer is already at the door; the door is being made the shape it expects.

5. The substrate

The agent economy sits on a model/dataset substrate that dwarfs the agent layer: 4.7M models, spaces, and datasets, linked by 2.87M dependency edges — the graph that will let a future edition name, when a server dies, exactly which agents are orphaned downstream.

Reproducibility

Individual data points are queryable point-in-time via the free public API (no key, 60 req/min): append ?as_of=YYYY-MM-DD to any /api/v1/entity/{owner}/{repo}. Aggregate datasets are CC-BY-4.0 and chain-anchored: Zenodo. Verify any dated observation at /verify; corrections at /dispute.

Caveats

404 = deleted or made private, not proven-gone-forever. MCP median lifespan is not reported (left-censored births). Migration Watch is a dated baseline, not a migration rate. The crawl figures are our own surface, our-traffic-excluded from 2026-08-28. Every population keeps its own source and method; method mixes are reported, never blended.