POIROT: Interrogating Agents for Failure Detection in Multi-Agent Systems

2026-06-01

AI Summary

Large language model (LLM)-driven multi-agent systems are vulnerable to subtle, emergent failures and hallucinations, which hinder their deployment in safety-critical applications. This work proposes a decentralized diagnostic protocol that leverages the cognitive diversity among agents to construct an intrinsic fault-detection layer, enabling self-diagnosis without external supervision through inter-agent interrogation and collective auditing. By circumventing the single-point failure limitations of centralized evaluation, the approach significantly outperforms single-LLM baselines across diverse complex tasks (OR = 1.60, p = 0.008), with performance gains amplifying as task complexity, agent count, and fault dimensionality increase. The authors further release POIROT, an open-source diagnostic library, alongside the BLAME benchmark to facilitate future research in autonomous multi-agent reliability.

Abstract

Orchestrating Large Language Models into Multi-Agent Systems (LLM-MAS) has unlocked remarkable reasoning capabilities, yet emergent failures and hallucinations that resist characterisation block their deployment in safety-critical domains -- a gap made legally untenable by emerging AI regulation. Existing evaluation paradigms share a common flaw: centralised judgment creates single points of failure and demands domain-specific expertise. Here we present POIROT, a protocol that repurposes a system's own agents as its diagnostic layer, leveraging the epistemic diversity already present in the architecture. Across evaluated settings, POIROT outperforms single-LLM evaluator baselines, with gains that scale with problem complexity (OR = 1.60, $p = 0.008$), agent count, and fault dimensionality, persisting under compound fault conditions. These results demonstrate that safety oversight need not be externalised: the agents executing a role carry sufficient collective intelligence to audit it. We release POIROT as an open-source library alongside BLAME, a benchmark for fault attribution in safety-critical multi-agent systems.

Problem

• failure detection
• multi-agent systems
• hallucination
• safety-critical
• LLM-MAS

Innovation

• multi-agent systems
• failure detection
• epistemic diversity
• self-auditing
• LLM orchestration