Living-Harness keeps agents from repeating the same failure

Living-Harness rewrites an AI agent's setup after each finished task so a fixed bug doesn't return, per a paper posted to arXiv.

The paper targets one failure: an agent fixes an error mid-task, then repeats it on a later task because nothing carried the fix forward. For teams running agents across repeated tasks, that gap is the cost, not the model's raw capability.

The fix works through what the paper calls an Evolution-SOP, a domain-level process for updating two things after each run. One is episodic memory: what triggered the failure, what the failure pattern looked like, and which recovery action worked. The other is a state graph, built from nodes, repair edges, and transition rules, that routes the agent around the same trap next time.

Tools and the agent's base context stay fixed throughout. Only the procedural layer changes: no bigger model, no new toolset, just accumulated repairs. The authors tested this on eight environments from τ²-Bench and MultiWOZ-2.4, established agent and dialogue benchmarks, rather than building a new one for the paper.