qa-agent-factory

P2P QA Agent Factory

Build QA agents in seven visible layers:

1. Mission/invariant — state what the agent is proving.
2. Narrow tools — expose only api_request(method, path, json_body) for HTTP.
3. Structured evidence — every tool call returns request, response, status, latency, interpretation slot.
4. Plan-act-observe-adapt prompt — Claude reasons about next steps from evidence.
5. Bounded self-healing — schema adaptation from validation errors, max 3 retries.
6. Judge/evals — verify state and business invariants, not status code only.
7. Regression handoff — each breach becomes a deterministic test.

Tie concepts to practice:

• Project rules keep generated code aligned with repo conventions.
• Reusable prompts prevent vague, one-off LLM behavior.
• Skills encode the workflow so we do not repeat mistakes.
• Agents isolate specialist reasoning: explorer, adversary, judge.
• Memory/context files give Claude stable project facts without dumping noise.
• Evals prove outputs match business rules.
• Guardrails limit tools and block dangerous operations.

Opening line:

I’ll build this in layers and run after each layer. Claude reasons over schemas and edge cases, but deterministic tools execute and log every action.