Ouroboros Self-Editing Agent Claims SOTA on Two Coding Benchmarks

A six-author paper out of a small research group is making an unusually direct claim: an agent that rewrites its own tools, prompts, context assembly, and core implementation can top two of the hardest agent benchmarks without a new base model doing the heavy lifting. The arXiv preprint describes Ouroboros as a self-developing coding agent whose improvements land as reviewed commits that then become the runtime for subsequent work.

The reported numbers are the reason to pay attention. On Terminal-Bench 2.1, an Opus 5 run scores 86.74%, described as the best result on the benchmark. On OSWorld-Verified, the same setup reaches 90.69%, above the prior best. A five-rollout CL-Bench campaign lands a normalized reward of 0.2301, which the authors also call state of the art. Evolution runs in two modes: recursive free evolution, where improvement itself is treated as a task, and experience-driven core evolution, where operational issues surface from ordinary work and trigger structural refinements. A companion deployment called Hope has been live for 161 days across seven communication surfaces, with the agent evolving through human interaction the whole time.

Before anyone re-plans a roadmap around this, it is worth being blunt about what the project page itself concedes. The Ouroboros site labels these as "self-reported results" with upstream review pending on some submissions, so the top-of-leaderboard framing has not been ratified by the benchmark maintainers. The paper also does not publish rejection rates from its three-stage plan/review/continue "immune system" during the 161-day Hope run, which is exactly where you would look to judge whether the human reviewers were doing real filtering or rubber-stamping.

Even with those hedges, the direction is what makes this a story. Roman Yampolskiy sits in the author list alongside the engineering team, which is a rare pairing of a safety-skeptical voice with a capabilities release. If the scores hold up under independent replication, the practical read for buyers and small teams is that agent quality can move meaningfully through disciplined harness engineering rather than only through the next model release, and the open-source posture means the harness itself is inspectable.