Living-Harness turns each completed agent trajectory into bounded updates to a persistent harness

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2607.26598 (July 29) targets the failure where an LLM agent recovers from an error within an episode but hits the identical failure in later tasks because post-episode feedback never revises the persistent harness. Guided by a domain-level Evolution-SOP, it writes two forms of procedural knowledge — episodic memory recording trigger conditions, failure patterns and recovery actions, plus a state graph of nodes, repair edges and transition rules — while freezing tools and base context so only procedural repairs accumulate. Evaluated on eight interactive environments derived from τ²-Bench and MultiWOZ-2.4 with improved average Pass@1.