Ouroboros is a self-developing agent harness where tools, prompts, and core implementation improve through reviewed commits that become the runtime for subsequent work. It operates via recursive free evolution or experience-driven core evolution triggered by bugs and inefficiencies found during use.
- On Terminal-Bench 2.1, an Opus 5 run scores 86.74%, the best reported result on the benchmark.
- On OSWorld-Verified, an Opus 5 run reaches 90.69%, exceeding the previous best score.
- A five-rollout CL-Bench campaign achieves a normalized reward of 0.2301, setting a new state of the art.
The system addresses operational safety by ensuring guardrails remain authoritative under evolutionary pressure, with benchmark campaigns using frozen snapshots while live evolution continues on separate lineages.