Living-Harness: Evolving Agent Frameworks by Turning Failures into Reusable Memory
burny_tech · x · 2026-07-31
LLM agents can recover from errors within a single task, but they often forget the fix once the episode ends. The paper Living-Harness Is an Interactive-Agent Evolver proposes a method to evolve the agent framework itself.
- Memory & State Graph: Converts evaluator feedback into episodic memory and a state graph, turning failures into reusable procedural repairs.
- Frozen Tools, Learning Workflow: Tools remain frozen, but the workflow learns when to trigger a repair, what action was missing, and how to recover next time.
- Performance: Achieves about +10 Pass@1 improvement over the strongest interactive baselines on the τ²-Bench and MultiWOZ-2.4 benchmarks.
More from coding & agent
- New ComfyUI Node for Krea 2 Fixes Multi-Character LoRA Face Blending — tekprodfx16 · 2026-07-31
- Generating a Complete FPS Game Demo with a Single Claude Prompt — FinanceYF5 · 2026-07-31
- GPT-5.6 Luna Tested After 80% Price Drop: Blazing Fast Full-Stack Code Generation — BorisMPower · 2026-07-31
- Run Local AI Agents on Consumer GPUs: Cloud Planning + Local Execution — fire_inabottle · 2026-07-31
- Devs Relate: Sometimes You Just Gotta Be Moral Support for Your AI Agent — mike64_t · 2026-07-31
- Google Releases Free 2-Hour Full Agent Engineering Course — goyalshaliniuk · 2026-07-31