ProofAgent-Harness stress-tests AI agent reliability

The framework rightly shifts focus from model capability to context engineering, acknowledging that agent failures are usually systemic context failures rather than LLM flaws.

* Shifting evaluation from simple benchmarks to multi-turn adversarial sessions is crucial for production agents.

* The multi-juror consensus model provides evidence-linked, auditable reports instead of generic success/failure flags.

* Out-of-the-box support for 183 bundled traps across 11 attack families makes it immediately useful for red-teaming and RAG verification.