Ouroboros: A Self-Rewriting Coding Agent Hits 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified — and Has Been Live-Evolving in Public for 161 Days

arXiv / HuggingFace Daily Papers (1,060 upvotes)·high signalXBlueskyLinkedInCopy link

arXiv 2608.08311 (submitted Aug 8, 2026) describes an agent that continuously rewrites its own tools, prompts, context assembly, and core implementation through reviewed commits, in two modes: recursive self-improvement cycles and experience-driven refinement from bugs it encounters. Running on Opus 5 it reports 86.74% on Terminal-Bench 2.1, 90.69% on OSWorld-Verified, and 0.2301 normalized reward on a 5-rollout CL-Bench campaign, claimed SOTA on each. The paper also documents 'Hope,' a 161-day public deployment evolving freely across seven surfaces under governed human communication — notable partly because Roman Yampolskiy, best known for arguing advanced AI is uncontrollable, is a co-author. It hit 1,060 upvotes on HuggingFace Daily Papers, roughly 13x the next paper that day.