<a href=\"https://github.com/microsoft/AutoSaddler\" rel=\"nofollow\">https://github.com/microsoft/AutoSaddler</a></p>\n","updatedAt":"2026-08-26T02:22:24.947Z","author":{"_id":"648804a275e78c1ebb7755bf","avatarUrl":"/avatars/d7f4fb2c07a4fc2dd72c5aa0b77d08da.svg","fullname":"Sungho Park","name":"pshlego","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6083397269248962},"editors":["pshlego"],"editorAvatarUrls":["/avatars/d7f4fb2c07a4fc2dd72c5aa0b77d08da.svg"],"reactions":[{"reaction":"👍","users":["obversarystudios"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.23041","authors":[{"_id":"6a8cfb688dd056518b7f543a","user":{"_id":"648804a275e78c1ebb7755bf","avatarUrl":"/avatars/d7f4fb2c07a4fc2dd72c5aa0b77d08da.svg","isPro":false,"fullname":"Sungho Park","user":"pshlego","type":"user","name":"pshlego"},"name":"Sungho Park","status":"claimed_verified","statusLastChangedAt":"2026-08-25T08:14:25.713Z","hidden":false},{"_id":"6a8cfb688dd056518b7f543b","user":{"_id":"67bc62b77727595ca5b6a4ca","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/JXilVx4ACCw6L9BnomqVP.png","isPro":false,"fullname":"Wonjoong Kim","user":"wjkim0229","type":"user","name":"wjkim0229"},"name":"Wonjoong Kim","status":"claimed_verified","statusLastChangedAt":"2026-08-25T16:10:54.527Z","hidden":false},{"_id":"6a8cfb688dd056518b7f543c","name":"Rongyuan Tan","hidden":false},{"_id":"6a8cfb688dd056518b7f543d","name":"Jue Zhang","hidden":false},{"_id":"6a8cfb688dd056518b7f543e","name":"Wook-Shin Han","hidden":false},{"_id":"6a8cfb688dd056518b7f543f","name":"Pengfei Gao","hidden":false},{"_id":"6a8cfb688dd056518b7f5440","name":"Chanyoung Park","hidden":false},{"_id":"6a8cfb688dd056518b7f5441","name":"Yongqiang Yao","hidden":false},{"_id":"6a8cfb688dd056518b7f5442","name":"Rao Fu","hidden":false},{"_id":"6a8cfb688dd056518b7f5443","name":"Elsie Nallipogu","hidden":false},{"_id":"6a8cfb688dd056518b7f5444","name":"Qingwei Lin","hidden":false},{"_id":"6a8cfb688dd056518b7f5445","name":"Saravan Rajmohan","hidden":false},{"_id":"6a8cfb688dd056518b7f5446","name":"Dongmei Zhang","hidden":false}],"publishedAt":"2026-08-24T00:00:00.000Z","submittedOnDailyAt":"2026-08-26T00:00:00.000Z","title":"AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces","submittedOnDailyBy":{"_id":"648804a275e78c1ebb7755bf","avatarUrl":"/avatars/d7f4fb2c07a4fc2dd72c5aa0b77d08da.svg","isPro":false,"fullname":"Sungho Park","user":"pshlego","type":"user","name":"pshlego"},"summary":"LLM agents remain unreliable on long-horizon tasks, where small local failures can compound over extended interactions and lead to overall task failure. Although external harnesses can substantially improve robustness, harness design remains a manual and expensive process that requires searching over a large space of prompts, tool configurations, and control logic. We propose AutoSaddler, an automatic harness optimization framework that formulates harness improvement as an offline learning problem and iteratively updates the harness using failure signals from mini-batches. AutoSaddler combines failure-trace diagnosis, structured patch generation that treats the harness as code, and validation-based update selection. Experiments on GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0 show that AutoSaddler substantially improves agent performance over the corresponding base harnesses, achieving gains of 9.0, 9.6, and 10.0 percentage points, respectively. Ablation studies further suggest that effective harness optimization benefits from three ingredients: deep debugging rather than shallow reflection, targeted modifications rather than unconstrained editing, and generalization-aware selection rather than trajectory-specific repair. Together, these results suggest that automatic harness optimization is a promising path toward more performant and reliable agent systems.","upvotes":43,"discussionId":"6a8cfb698dd056518b7f5447","projectPage":"https://autosaddler-projectpage.github.io/","githubRepo":"https://github.com/microsoft/AutoSaddler","githubRepoAddedBy":"user","ai_summary":"AutoSaddler automatically improves LLM agent harnesses via offline failure-driven optimization, boosting performance on long-horizon benchmarks.","ai_keywords":["LLM agents","harness optimization","failure-trace diagnosis","structured patch generation","validation-based update selection","deep debugging","targeted modifications","generalization-aware selection"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":6,"organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"648804a275e78c1ebb7755bf","avatarUrl":"/avatars/d7f4fb2c07a4fc2dd72c5aa0b77d08da.svg","isPro":false,"fullname":"Sungho Park","user":"pshlego","type":"user"},{"_id":"6574932cd0ed8f5761069ac2","avatarUrl":"/avatars/2a6b6b9e8b01179eacc92d2a9cbf5df9.svg","isPro":true,"fullname":"Seungtae","user":"stnamjef","type":"user"},{"_id":"67bc62b77727595ca5b6a4ca","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/JXilVx4ACCw6L9BnomqVP.png","isPro":false,"fullname":"Wonjoong Kim","user":"wjkim0229","type":"user"},{"_id":"678f172a02a1c58b5b3086f6","avatarUrl":"/avatars/ed72e9aacbf75f73406317d0adc00771.svg","isPro":false,"fullname":"Sunho Cha","user":"carprefer","type":"user"},{"_id":"6858f58b48fab0227165bca3","avatarUrl":"/avatars/ac0e7a3d2c225f50c16f7bca68a8ca82.svg","isPro":false,"fullname":"HyoJeong Yun","user":"hyojeongyun","type":"user"},{"_id":"64c52ee547418a0a59b3d9b3","avatarUrl":"/avatars/a1ef01638b1cbbdaf5b5328d3dbdae90.svg","isPro":false,"fullname":"Wook-Shin Han","user":"wscrony","type":"user"},{"_id":"63840be059424581c3611bf3","avatarUrl":"/avatars/e2c32fac921a5305a5f89c3c86f690ef.svg","isPro":false,"fullname":"hello","user":"dododo71","type":"user"},{"_id":"60af909e288a0f96f6cefc4d","avatarUrl":"/avatars/44a8ab48acb58c45de0b0947a1b56e7c.svg","isPro":false,"fullname":"Yeonjun In","user":"Yeonjun","type":"user"},{"_id":"65f40e43653c231cbaf7d1e4","avatarUrl":"/avatars/a42ac5454cbe175f04c3420fce90cad2.svg","isPro":false,"fullname":"Jue Zhang","user":"JueZhang","type":"user"},{"_id":"668215234905815dcf6bc754","avatarUrl":"/avatars/0df14f7674a056cb0928a0cdf89ae4b0.svg","isPro":false,"fullname":"Hyunchul Kim","user":"khchul","type":"user"},{"_id":"65d595a7b4c1f26f30206644","avatarUrl":"/avatars/e3550742181b5a85ce93803fd7151fd3.svg","isPro":false,"fullname":"jdsj","user":"jdsj","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":3,"organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.23041.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2608.23041 in a model README.md to link it from this page.

Cite arxiv.org/abs/2608.23041 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2608.23041 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.