Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
About

A public, relational corpus of the AI agent ecosystem

Every entity, whether an agent, a foundation model, a memory provider, a protocol or a security incident, is drawn from a primary source and joined into a single knowledge graph. The value is not the rows, it is the map between them.

Explore the corpus
The maintenance ecosystem

The guide about AI agents is, by design, kept current by a small ecosystem of them. Its shape is a blackboard: three agents work asynchronously and coordinate through one shared corpus, never peer-to-peer. Every proposed change filters down to a human gate before it reaches the site. Each agent is itself a topology this guide catalogues.

The three agents · by temperament

Split by temperament and trust boundary, not bundled. Discovery wants recall; curation wants precision; a public chat must stay read-only. One agent cannot be all three at once, so it is three, and each is itself a morphology this guide catalogues.

Atlas, map-reduce morphology
Atlas
Research · scout

The outward-facing agent. Fans out across the open web and the literature to find new agents, sources and topologies, then reduces the noise to a short list of sourced candidates.

temperamentoutward · high-recall
cadenceweekly scout · quarterly source scan
toolsweb search · fetch · credibility filter · MCP client
readsthe open web + the corpus
writescandidates only, handed to the Steward via MCP
the Steward, feedback-gate morphology
the Steward
Maintenance · curator

The inward-facing agent. Verifies each candidate against the credibility filter, keeps existing rows fresh, demotes or prunes what no longer holds, and opens a pull request. Never publishes unaided.

temperamentinward · high-precision
cadenceon candidate handoff · weekly review-due pass
toolscredibility filter · build.mjs · GitHub PR API
readsthe corpus + Atlas’s candidates
writesa reviewable PR, the only write path
the Concierge, context-injection morphology
the Concierge
Chat · read-only

The public-facing agent. Answers questions grounded strictly in the corpus, cites every claim, and says "unknown" rather than guessing. Read-only, so untrusted input can never reach the data.

temperamentpublic · read-only
cadenceper query
toolscorpus retrieval · citation gate
readsthe corpus
writesnothing
Status · building in the open

The ecosystem is in development. This page describes the architecture we are building; it does not claim a live run. When the agents are running, this page will show the last maintenance pass, how many sources were re-verified and when. Until then every change to the guide is made and reviewed by hand, on the same propose-then-approve gate the Steward will inherit. Nothing is published to the guide without a human signing off, agent or not.