Found first: a primary source the press has not covered yet.
A new arxiv paper by Anton Razzhigaev and colleagues introduces Ouroboros, an agent that rewrites its own code, tools, and prompts through reviewed commits, then runs subsequent tasks on the updated system. The paper reports 86.74% on Terminal-Bench 2.1, which the authors describe as the best result reported on that benchmark, and 90.69% on OSWorld-Verified, exceeding the best previously reported score on that evaluation. Both results were achieved with Claude Opus 5 as the underlying model.
What the source says
Ouroboros evolves through two mechanisms: recursive free evolution, where self-improvement is itself a schedulable task and one cycle can trigger the next, and experience-driven core evolution, where bugs and inefficiencies encountered during real work prompt reviewed structural changes to the agent. The authors have run a deployment called "Hope" continuously for 161 days, described in the paper as a living agent experiment in free evolution under governed human communication across seven surfaces. On CL-Bench, using a five-rollout campaign, the agent achieved a normalized reward of 0.2301, which the authors describe as a new state of the art. The paper's safety section frames keeping guardrails authoritative under evolutionary pressure and public social interaction as a primary design problem.
Why it matters
Agent benchmark progress has largely tracked model releases. This paper describes a system that reached the top of two hard benchmarks by modifying its own implementation; the underlying model stayed constant. A deployment called "Hope" has been running continuously for 161 days. For practitioners building production agents, the safety framing is the most directly applicable part: the paper addresses how to keep guardrails stable when the agent can rewrite the code those guardrails live in.