{"id":"arxiv:2606.01770","title":"Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams","date":"2026-06-01","section":"Harness","quality":0.233234,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/","source_url":"https://arxiv.org/abs/2606.01770","text_available":true,"authors":["Zewen Liu","Zhan Shi","Yisi Sang","Bing He"],"outgoing":[{"source":"arxiv:2606.01770","target":"arxiv:2507.19457","evidence":"id","id":"arxiv:2507.19457","title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/"},{"source":"arxiv:2606.01770","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2603.28052","title":"Meta-Harness: End-to-End Optimization of Model Harnesses","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/"}],"incoming":[],"text_parts":[{"start":0,"end":15328,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html"},{"start":15328,"end":31284,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html"},{"start":31284,"end":46948,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html"},{"start":46948,"end":50453,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-4.html"}],"contents":[{"title":"Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"Abstract","offset":110,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"1 Introduction","offset":1418,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"2 Related Work","offset":6532,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3 Method","offset":8063,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3.1 Problem Formulation","offset":8076,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3.2 Analytical Framework","offset":9173,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"Proposition 1 (Regret Decomposition) .","offset":11574,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3.3 Sustained Auto-Harness via Multi-Agent Evolution","offset":13329,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3.4 Solve-Time Adaptation via Harness-Tree Routing","offset":15272,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-1.html#text"},{"title":"3.5 Human-in-the-Loop Channel","offset":17474,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4 Experiments","offset":18677,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.1 Experimental Setup","offset":18695,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.2 Comparison with Baselines","offset":21178,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.3 Benchmark Bottlenecks","offset":22399,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.4 Stateful Multi-Agent Evolution","offset":25133,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.5 Solve-Time Routing on the Harness Tree","offset":26197,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"4.6 Human Steering for Auto-Harnessing","offset":27177,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"5 Conclusions","offset":28145,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"6 Limitations","offset":28997,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"7 Ethics Statement","offset":29747,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"8 AI Usage Statement","offset":30092,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"Appendix A Benchmark and Evaluation Details","offset":30255,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"A.1 PolyBench","offset":30803,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-2.html#text"},{"title":"A.2 CTF-Dojo","offset":32755,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"A.3 FutureX","offset":33486,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix B Implementation and Reproducibility Details","offset":35012,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix C Benchmark Non-Stationarity Diagnostics","offset":38871,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix D Further Experiments","offset":41885,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix E Per-Domain and Per-Category Breakdowns","offset":42618,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix F Multi-Agent Evolution Dynamics","offset":43762,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix G Routing Behaviour and Branch Performance","offset":44630,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-3.html#text"},{"title":"Appendix H Human-in-the-Loop Event Log","offset":47492,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-4.html#text"},{"title":"Appendix I Run-Detail Analysis","offset":48421,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-4.html#text"},{"title":"Appendix J System Prompts","offset":49105,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-4.html#text"},{"title":"Instructions for reporting errors","offset":49489,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/part-4.html#text"}],"text_code_points":50453,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/source.md"}