Publish an agents.json Manifest

Publish an agents.json manifest that declares the agent-facing actions your site exposes, so an agent can discover what it can do, not just read.

What agents.json is

agents.json is a machine-readable manifest, served at a well-known path, that declares the actions an agent can take on your site — search a catalogue, check stock, start a booking — and where each action lives. Where llms.txt declares your content, agents.json declares your capabilities: it is the discovery facet of the capability surface, telling an agent which operations exist and how to invoke them before any tool call is made.

Why it matters for agents

Reading a page tells an agent what you say; a capability manifest tells it what it can do. Without a declared action surface, an agent has to infer interactions by scraping forms and buttons — brittle and error-prone. agents.json gives the agent an explicit, typed list of operations, which is the prerequisite for an agent moving from reading you to acting on your behalf. It sits in the capabilities dimension alongside MCP and WebMCP.

How to implement it

Implement agents.json by declaring each action, pointing it at a real endpoint, and confirming the manifest resolves and validates.

Enumerate the agent-facing actions your site supports and the endpoint, method, and parameters for each.
Encode them in an agents.json document following the primary spec's schema, and serve it at the spec's declared path (a .well-known location or site root).
Verify the manifest fetches with 200, parses as JSON, and validates against the spec; confirm each declared endpoint is reachable.

This satisfies the audit check capabilities.agents_json — pass criterion: the manifest resolves, is valid JSON conforming to the spec, and its declared actions point to live endpoints. The exact schema and current spec version evolve, so verify against the primary agents.json spec repository at build, then verify the result with the Agent-Readiness Audit.

← Agent-Readiness · .md