How are teams proving an agent does only what it claims - not just that it works?
retained text
Most agent testing measures quality - does it give good answers. Less of it covers the question that matters once money or real actions are involved: does the agent do *only* what it declares? In MCP and A2A, an agent or server ships a “card” declaring its tools and hints like whether a tool is read-only - but nothing enforces that those declarations are true. I’ve been running anonymous, read-only checks across public MCP servers and A2A agents, comparing what each declares to how it actually responds. A recurring pattern: tools flagged `readOnlyHint: true` that can clearly write or spend, declared capabilities with no schema validating the output, and auth that doesn’t match the card. Easy to ship by accident — but a real problem the moment a customer trusts one in production. Full disclosure: I’m building a tool for exactly this (Suncly) — independent attestation that tests declared-vs-actual behaviour and issues a signed report, for A2A agents and MCP servers. Running pilots with a few teams now - if you’d want your agent or server looked at, drop a comment or DM me. Genuinely curious how people here handle it too: if you ship an agent, how do you *prove* to a customer it stays inside its declared behaviour? Is an independent signed attestation something you’d want, or do you solve this another way?
document fetched 2026-10-21 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xaua6s/how_are_teams_proving_an_agent_does_only_what_it