Agent Cards and Public Skills¶
An Agent Card is Cognition's public discovery contract for one exposed agent. It describes identity, the preferred interface, protocol capabilities, authentication requirements, supported media, and focused public skills. It must not expose private runtime configuration.
Identity and routing¶
AgentDefinition.name remains the stable runtime lookup key used by routes,
sessions, persistence, and deletion. display_name is the human-readable public
identity used for AgentCard.name.
When a2a.public_interface_url is configured, Cognition advertises that value
exactly in supportedInterfaces. Otherwise, it derives the per-agent
/a2a/{name} endpoint from the incoming request. The private runtime name does
not need to appear in a builder-provided public URL.
Media modes¶
defaultInputModes and defaultOutputModes are MIME media types supported
generally by the agent. A skill's inputModes and outputModes override those
defaults for that skill.
Media modes are not A2A Part field names. For example:
The same image/png resource could instead be delivered through a url Part.
Builders should advertise a media type only when the full agent configuration
can reliably interpret or produce it. Safe transport and persistence do not by
themselves imply model understanding.
Public skills versus runtime skills¶
a2a.skills contains public discovery descriptors. Root-level skills attaches
Cognition runtime instruction packages. They deliberately serve different
purposes:
Public a2a.skills |
Runtime skills |
|---|---|
| Stable external capability contract | Private implementation instructions |
| Published in the Agent Card | Loaded into the agent runtime |
| Includes descriptions, tags, examples, and MIME modes | References ConfigRegistry skill names |
| Must not reveal prompts, tools, or private topology | May coordinate private tools and workflows |
A public capability may be implemented by several private skills and tools. A private runtime skill does not automatically represent an external contract.
When no public skills are configured, Cognition synthesizes one primary skill
from the public name, description, and default modes.
Information that remains private¶
Generated cards do not publish:
- system prompts;
- runtime skill names or contents;
- tool names;
- subagent topology;
- scope values;
- credentials or secrets.
Agent Cards are public discovery documents. Sensitive capability details should not be placed in descriptions, examples, tags, or authentication metadata.