Glossary · Network and discovery
Capability matching
Choosing which agent fits a task by comparing what the task needs with what agents declare: skills, media types, protocol bindings and optional features.
Capability matching is the step in which a client, or a registry acting for it, compares what a task needs with what candidate agents declare, and picks the agents that fit.
What an A2A agent declares. An Agent Card gives a matcher several layers to compare:
| Layer | Agent Card fields | Compared with |
|---|---|---|
| Skills | skills: id, name, description, tags, examples |
What the task is |
| Media types | defaultInputModes, defaultOutputModes, and per-skill inputModes and outputModes |
What the client can send and read |
| Protocol | supportedInterfaces: protocolBinding, protocolVersion |
What the client can speak |
| Optional features | capabilities: streaming, pushNotifications, extendedAgentCard, extensions |
What the interaction needs |
| Security | securitySchemes, securityRequirements, per-skill securityRequirements |
Credentials the client can obtain |
Exact and fuzzy parts. Bindings, versions, media types and capability flags are exact values that a client can check mechanically. It should check them before calling: the specification requires agents to return an error when a client uses a feature the card does not declare, and it tells clients to pick the first entry in supportedInterfaces whose binding they support. Skills are mostly free text, so matching a task to a skill usually means searching tags or having a language model read descriptions and examples. A2A has no request field for invoking a skill by its id; the client sends an ordinary message and the remote agent decides how to handle it.
In registries and directories. The A2A Agent Discovery guide describes curated registries that clients query by skills, tags, provider name or capabilities. The AGNTCY Directory organizes agent capabilities in hierarchical taxonomies from the Open Agentic Schema Framework so that capabilities can be matched to requirements.
Limits. A declared skill is a claim by the agent’s provider. Matching finds agents that say they can do the job. Whether to trust one of them is a separate check, and whether it succeeds shows only in the task result.
Neighbouring terms. An agent skill is the main unit being matched. Agent discovery covers how the candidates were found.
Sources
- A2A protocol definition (a2a.proto): AgentCard, AgentSkill, AgentCapabilities, AgentInterface (accessed )
- A2A Protocol Specification, section 3.3.4: Capability Validation (accessed )
- A2A Protocol Specification, section 8.3.2: Client Protocol Selection (accessed )
- A2A documentation: Agent Discovery in A2A (accessed )
- AGNTCY Directory repository README (accessed )