Living-Harness: Self-Evolving LLM Agent Framework

Researchers propose Living-Harness, a self-evolving agent harness for large language models (LLMs) that addresses the problem of recurring execution failures. Traditional static harnesses improve reliability through fixed tools and workflows but remain unchanged after deployment, allowing the same failures to reappear. Living-Harness converts each completed trajectory and evaluator signals into posterior evidence for bounded harness updates, guided by a domain-level Evolution-SOP (Standard Operating Procedure). It extracts episode abstractions and structured update evidence, writing two forms of procedural knowledge: episodic memory (recording trigger conditions, failure patterns, and recovery actions) and a state graph. The paper is published on arXiv under ID 2607.26598.