The research behind ProofAgent — published papers on AI agent evaluation

ProofAgent is grounded in five published papers by Fouad Bousetouane on adversarial AI agent evaluation, context engineering, memory control, and production readiness. All are open methodology and reproducible from open source code.

ProofAgent Harness: Adversarial Evaluation for Production AI Agents (arXiv:2605.24134)

The foundational whitepaper. A full pipeline of planning, adversarial conducting, multi juror scoring, debate consensus, and signed reporting, with a 183 trap library across 11 attack families, composite attack chains, and a six metric rubric. The headline result: production grade agents on frontier LLMs fail under sustained adversarial pressure, so the agent layer needs its own stress testing infrastructure.

Human-on-the-Bridge: Scalable Evaluation for AI Agents (arXiv:2606.16871)

The paradigm behind the Harness. Instead of a human reviewer in the loop on every run, curate human expertise once, upstream, into reusable traps, juror personas, and rubrics, then let small evaluator models stress test frontier class agents at scale. Rigorous because it carries real expertise, scalable because it runs automatically, locally, and cheaply.

AI Agents Do Not Fail Alone: The Context Fails First (arXiv:2607.14275)

The empirical validation of context engineering as a leading indicator of agent reliability. Holding frontier agents fixed and varying only their context, context quality criteria predict behavior: grounding predicts hallucination resistance, guardrail coverage predicts manipulation resistance.

AI Agents Need Memory Control Over More Context (arXiv:2601.11653)

Why long running agents drift. Transcript replay and retrieval based memory grow context without bound, accumulate stale errors, and pull multi turn decisions off their constraints. The paper argues for governed memory control, deciding what an agent retains, over ever larger context windows.

Stop Shipping AI Agents on Faith: Capability Is Not Production Readiness (arXiv:2607.27677)

The paper that defines PAI, the ProofAgent Index: a production readiness index over four axes (evaluation, context, compliance, governance), computed on every harness run and gateable in CI. Capability tells you what an agent can do; PAI tells you whether it is ready.

Open methodology, full reproducibility

All evaluation methodology, trap libraries, juror personas, and scoring rubrics ship open source under Apache 2.0 in the ProofAgent Harness. Every result in the papers is reproducible from the published code with any LiteLLM supported model.