A new framework watches AI agent networks for teammates that go quietly wrong, not just ones that go offline.
Researchers built MeshHeal, a fully decentralized self-healing framework for networks of LLM agents, to catch gray failures: agents that keep responding normally while the quality of their output quietly degrades. The system works on two timescales: a fast layer escalates shaky or low-scoring outputs from a single reviewer to a full committee and corrects them before use, while a slower layer compares an agent's scores against peers on similar tasks to separate real, persistent decline from normal noise, then benches the agent from routing until recovery probes show it has improved. The team also built a new test, Model-Backed MAS Evaluation, because judging agent ability from prompts alone can hide routing mistakes. On the BBH, MATH, and MMLU-Pro benchmarks, MeshHeal hit 0.839 accuracy during simulated degradation using 51,000 tokens per task, versus 0.807 accuracy at 115,000 tokens per task for the strongest baseline, Symphony.
Multi-agent LLM systems are starting to handle real workloads, and the failure mode most designs plan for, a crashed or unresponsive agent, is the easy one. An agent that stays online but is subtly wrong is the one that quietly corrupts downstream tasks, and MeshHeal's real pitch is efficiency: it beat its rival's accuracy using less than half the tokens.
That said, this is one paper's results against a single rival system on curated benchmarks, and it says nothing about how the approach holds up against agents that are actively adversarial rather than just quietly degraded.