Feed
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
A coding agent rewrites its own harness through reviewed commits, then tops agent benchmarks.
86.74%%
Terminal-Bench 2.1 audited accuracy
Best reported score on 89 hard terminal tasks
How it works
Harnesses stay frozen
Agent scores depend on how the harness builds context and uses tools, yet most harnesses never change after design.
Agent edits its harness
The agent improves its own tools, prompts and code. Each change passes multi-model diff review before it becomes live.
Top benchmark scores
Frozen snapshots lead Terminal-Bench 2.1, OSWorld-Verified and CL-Bench, and match Codex on SWE-bench Pro.
From the paper