Agent Card: Agent vs System vs Human Readable Annotations

**Question: How should Agent Cards indicate machine-readable vs. agent-visible semantics?**

Is there an emerging view in A2A on how an Agent Card should distinguish between fields intended for machine-readable discovery/orchestration and fields intended to be surfaced to an agent or LLM?

Today, A2A does not appear to have annotation semantics comparable to MCP, such as `readOnly`, `destructive` or other capability/usage hints. This may be suitable for an extension, but the broader question remains: who or what is expected to perform discovery across an Agent Card, and which parts of the card should be surfaced to which consumers?

Without clear annotations, it is difficult to translate an Agent Card and its skills into provider-independent LLM calls. As far as I can tell, no SDK today can reliably perform that translation from an Agent Card alone, which feels like a gap.

Looking to open a discussion on this and if there are semantic patterns that should be embedded into the protocol usage.

This is a design discussion about semantic annotations in the A2A protocol's Agent Cards. Start by reading the A2A specification and existing Agent Card examples to understand the current structure. Look for prior discussions on protocol extensions or comparable annotations in MCP. The goal is to propose a pattern, not to implement a specific change.

Vom Indexierungsmodell aus dem Issue-Text verfasst.