Taming "Zombie'' Agents: A Markov State-Aware Framework for Resilient Multi-Agent Evolution

In three linesAgentRevive introduces a Markov state-aware framework for resilient multi-agent LLM system evolution. Instead of aggressively pruning failing agents, the method uses soft state transitions (Active/Standby/Terminated) with a hallucination risk estimator. Results: outperforms baselines on general reasoning, domain-specific tasks, and hallucination challenges while reducing token consumption.

Summary generated by Claude — human-verified