Living-Harness lets LLM agents self-evolve, boosting success by 10%

Large language model (LLM) agents typically rely on static harnesses—fixed tools, context, memory, and workflow structures—that remain unchanged after deployment. This means the same failure can reappear across different tasks because post-episode feedback rarely revises the underlying guidance system. To solve this, researchers from multiple institutions introduce Living-Harness, a framework that allows the agent harness to evolve autonomously. After each completed trajectory, Living-Harness extracts episode abstractions and structured update signals guided by a domain-level Evolution-SOP (Standard Operating Procedure). It then writes two forms of procedural knowledge: episodic memory capturing trigger conditions, failure patterns, and recovery actions, and a state graph storing state nodes, repair edges, and transition rules. The updated harness is retrieved for future interactions, while tools and base context remain frozen. This design ensures that procedural fixes accumulate across evolution cycles without retraining the underlying LLM.

In experiments across eight interactive environments derived from τ²-Bench and MultiWOZ-2.4, Living-Harness boosted average Pass@1 by 10.07 and 9.91 percentage points respectively over the strongest interactive baselines. The evolved harness state can be reused across different model backbones via retrieval-only operations, making it model-agnostic and efficient. This represents a shift from static agent frameworks to adaptive, self-improving systems that learn from their own execution history. For developers building reliable AI agents, Living-Harness offers a practical way to reduce repetitive errors and improve task completion rates without manual intervention or expensive fine-tuning.