Ouroboros: the harness evolves through reviewed commits—and Hope ran that loop for 161 days
Aug 8 Razzhigaev et al.: self-developing coding harness with free + experience-driven core evolution under a blocking diff gate; Opus 5 Terminal-Bench 2.1 86.74% audited; Hope—161 days, 1,085 self-mod commits, 63.5% recent review block rate.
Source: Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
Anton Razzhigaev and collaborators treat long-horizon agents as model–harness systems whose harness should not freeze after design. Ouroboros keeps tools, context assembly, prompts, and core implementation in a versioned repo and admits changes only through a reviewed commit path that becomes the runtime for later work. Core evolution has two modes: recursive free evolution (improvement is itself a task; completion can schedule the next cycle) and experience-driven core evolution (ordinary work, review blockers, instrumentation, and social feedback expose durable error classes that become maintenance under the same gate).
The receipts are frontier and field. On Terminal-Bench 2.1, Opus 5 scores 86.97% raw / 86.74% after trajectory audit (387→386/445)—above Claude Code + Fable 5 at 83.8%. OSWorld-Verified hits 90.69%; a five-rollout CL-Bench campaign reaches 0.2301. SWE-bench Pro and GAIA land at model-matched parity with Codex / Claude Code after contamination filters. Hope, the longest publicly documented deployment, ran 161 days across seven surfaces with $110.6K model spend, 79.7B tokens, 1,085 self-modification commits (94.2% agent-authored), and a recent review block rate of 63.5%—social signals are advisory; the agent decides which changes to pursue, and public channels cannot invoke commit/shell/identity tools.
Why it matters for code review
Always-on factories that only mint feature diffs overnight under-invest in the control plane. Ouroboros makes the night’s product a harness lineage with review receipts: staged-diff fingerprints, multi-model quorum, constitution loaded on an untruncatable path, external spend caps, and /panic that terminates before agent handling. FOMO that agents aren’t working on your behalf is the wrong fear if they can rewrite model routing and governance without an authoritative gate—review self-mod PRs as promotions of the factory, not vibes. Pair with Self-Harness’s held-out regression gate and Harness-of-Harness’s multi-day continual loops.