verify

verify — delegate a verification check

Ask a peer agent to evaluate an artifact (code, a plan, a doc, a diff) against explicit criteria and return a structured verdict. This is the Anthropic-validated pattern that's consistently worth the extra tokens: a second agent either confirms the work or flags a specific problem — it does not rewrite.

When to use

• You produced something and want an independent pass/fail, not a revision.
• The criteria are enumerable ("does X, handles edge case Y, passes test Z").
• You want the verifier's disagreement surfaced as evidence, not silently merged back into your own output.

When NOT to use

• You want the peer to do the work. That's delegation, not verification — prompt a peer directly without this pattern.
• The criteria are vague ("is this good?"). Verifiers cannot return useful structured verdicts without concrete checks.
• The cost of the second call is not justified. Verification typically 2–3× the tokens of the original turn.

Protocol

On the A2A inbound surface, set Message.metadata.return_format to "verdict" when sending message/stream:

{
  "message": {
    "parts": [{ "kind": "text", "text": "<verifier prompt>" }],
    "metadata": { "return_format": "verdict" }
  }
}

The peer responds with an A2A artifact-update whose part has kind: "data" and mimeType: "application/vnd.a2a-bridge.verdict+json". The data object matches the verification artifact shape from docs/design/architecture.md §"Verification artifact":

{
  "verdict": "pass" | "fail" | "needs-info",
  "reasoning": "one to three sentences",
  "evidence": [
    { "claim": "string", "source": "file:line | url | inline", "note": "optional" }
  ],
  "followups": ["string", ...]
}

Unrecognized verdict values are coerced to "needs-info". An empty or missing reasoning is rejected — verifiers that cannot articulate reasoning must return "needs-info" with a short note instead.

Prompt scaffold