Before one agent can delegate to another, it needs to know what the other can do and how to reach it. A2A handles this with Agent Cards.
What an Agent Card Contains
- Name, description and provider.
- The service endpoint URL.
- Supported capabilities, such as streaming and push notifications.
- Authentication requirements.
- Skills: what the agent can do, with descriptions, examples, and accepted input and output types.
Discovery Approaches
- Well-known location: agents publish their card at a standard path on their domain.
- Registries or catalogues: organisations maintain directories of approved agents.
- Direct configuration: client agents are configured with specific cards.
Choosing an Agent
A client agent — or the model inside it — reads skill descriptions to decide which remote agent suits a task, much as a model reads tool descriptions.
Writing Good Cards
Describe skills precisely, with clear examples and limits. Vague cards lead to wrong delegation.
Security
Cards can be signed or served over authenticated channels in some deployments. Treat discovered cards from untrusted sources cautiously, and keep sensitive capabilities in authenticated extended cards.