POIROT Makes AI Agents Catch Each Other's Mistakes 60% More Accurately
POIROT turns a multi-agent system's own agents into its safety auditors, outperforming single-LLM evaluators with gains that scale with complexity
Takeaways
- POIROT is a new protocol that uses a multi-agent system's own agents as its fault detectors, replacing external single-LLM judges.
- It outperforms single-LLM evaluator baselines with an odds ratio of 1.60 (p=0.008), and gains scale with system complexity and agent count.
- The five-phase protocol includes hazard space definition, self-assessment, peer interrogation, private voting, and weighted aggregation.
- Performance holds under compound fault conditions where multiple agents fail simultaneously -- the hardest real-world scenario.
- The team also releases BLAME, a new open-source benchmark for fault attribution in safety-critical multi-agent systems.
- Prior SOTA methods top out at 53.5% accuracy on failure attribution; POIROT's decentralized approach significantly improves on this baseline.
Deploying LLM-based multi-agent systems in high-stakes environments -- think clinical decision support, autonomous trading, or legal document review -- has been stuck behind a stubborn wall: nobody has a reliable, scalable way to catch when one of those agents goes wrong. A new paper introduces POIROT, a diagnostic protocol that sidesteps the need for an external watchdog entirely by turning the system's own agents into its fault detectors.
The name is a nod to Agatha Christie's famous detective, and the analogy is intentional. Rather than calling in an outside investigator, POIROT interrogates the suspects already in the room.
The Problem with Watching from the Outside
Large Language Model-based Multi-Agent Systems (LLM-MAS) have emerged as flexible reasoning engines capable of tackling complex tasks through the coordinated action of specialised agents. Their adoption in non-critical commercial applications -- from automated customer support and software engineering assistants to document processing pipelines -- is already widespread. But their deployment in safety-critical domains remains largely out of reach.
The standard approach to catching failures is the LLM-as-a-judge paradigm: a single, powerful model (say, GPT-4) is given the system's outputs and asked to score them. The problem is that this creates a single point of failure with well-documented biases. Early LLM-as-judge methods typically involve a single model passing judgment, which introduces its own limitations -- a single LLM judge may carry inherent biases, for instance preferring certain writing styles or content, and thus produce skewed evaluations. They can be gamed by adversarial outputs, and models also tend to favor outputs that resemble their own phrasing or length -- a bias towards verbosity and familiar styles.
The deeper issue is structural. Existing evaluation paradigms share a common flaw: centralised judgment creates single points of failure and demands domain-specific expertise. In complex, multi-step pipelines, a single external judge also faces