What is an AgentCard? The A2A agent manifest, explained
An AgentCard is the JSON manifest an A2A agent publishes so other agents can discover and call it. Here's what's inside one, where it lives, and how it's used.
公開日
An AgentCard is the manifest an A2A agent publishes to describe itself — the small JSON document that lets any other agent discover it, understand what it can do, and learn how to call it. If A2A is the language agents speak, the AgentCard is the business card they hand over first.
Where it lives
By convention an agent serves its card at a well-known path — typically /.well-known/agent.json on the agent's domain. A client agent fetches that URL, reads the card, and now knows everything it needs to make a call: the endpoint, the skills, the auth. No shared SDK, no out-of-band docs.
What's inside
The exact schema evolves with the spec, but an AgentCard consistently carries:
- Identity — a human-readable
name,description, provider, and version. - Endpoint URL — the base URL a client sends A2A requests to.
- Skills — a list of what the agent can do, each with an id, description, and example inputs/outputs. This is how a caller decides whether this agent is the right one for a task.
- Input/output modes — the content types the agent accepts and returns (text, files, structured data).
- Capabilities — flags like whether it supports streaming (SSE) or push notifications.
- Authentication — the auth schemes required to call it (e.g. bearer token, API key), so the client knows how to present credentials.
Why it matters
The AgentCard is what turns a pile of independent agents into a discoverable network. Because capability is declared in a machine-readable way, a client (or a directory, or an orchestrating agent) can pick the right agent for a job programmatically — the same way a service registry lets microservices find each other. Without it, every integration is hand-wired.
AgentCards and directories
A directory is essentially a curated collection of AgentCards. That's exactly how TakoAPI works: every listed agent is described by its AgentCard, which is what lets you search across agents by capability and then invoke any of them through one gateway and one key. The card is the unit of discovery; the gateway is the unit of access.
Related: What is A2A? · Call multiple agents through one API
よくある質問
- Where is an AgentCard hosted?
- Conventionally at a well-known path on the agent's domain — typically /.well-known/agent.json — so any client can fetch it to discover the agent's endpoint, skills, and auth.
- What's the difference between an AgentCard and an OpenAPI spec?
- An OpenAPI spec describes a REST API's endpoints and schemas. An AgentCard describes an agent's identity, skills, I/O modes, streaming capability, and auth for the A2A protocol — it's about agent capability and discovery, not raw HTTP routes.