Ouroboros rewrites its own core, and it just took Terminal-Bench
A coding agent that edits its own source code, gets the diff reviewed, commits it, and then keeps working. That's Ouroboros, arXiv 2608.08311, out August 8 and now sitting at over a thousand upvotes on Hugging Face Daily Papers — the highest number we've seen on an agent paper all year.
The numbers are what make it hard to wave off. Terminal-Bench 2.1 at 86.74 percent on Opus 5. OSWorld-Verified at 90.69 percent. CL-Bench normalized reward 0.2301. These are frontier-tier results from a system whose main claim is that nobody hand-tuned the scaffold — the scaffold tuned itself.
Two evolution modes. Recursive free evolution, where "improve yourself" is just another item on the scheduler. And experience-driven evolution, where a bug or a slow path hit during real work becomes the trigger for a structural change. Tools, prompts, context assembly, core implementation — all of it is in scope, all of it goes through a code review gate before landing. That gate is the whole design. We covered VaG two days ago making exactly this argument: an agent generating its own improvements is fine, an agent approving them is not.
The detail buried in the paper is the one worth remembering. A deployment called Hope has been running 161 days in free evolution across seven surfaces, under governed human communication. Not a benchmark run — a living agent that has been rewriting itself for five months. The authors are explicit that safety is the binding constraint here, since a self-modifying agent that can also pick which model API it calls is a different animal than one that can only edit its prompts. Roman Yampolskiy is on the author list, which tells you they know it.
Paper: https://arxiv.org/abs/2608.08311
← Back to all articles
The numbers are what make it hard to wave off. Terminal-Bench 2.1 at 86.74 percent on Opus 5. OSWorld-Verified at 90.69 percent. CL-Bench normalized reward 0.2301. These are frontier-tier results from a system whose main claim is that nobody hand-tuned the scaffold — the scaffold tuned itself.
Two evolution modes. Recursive free evolution, where "improve yourself" is just another item on the scheduler. And experience-driven evolution, where a bug or a slow path hit during real work becomes the trigger for a structural change. Tools, prompts, context assembly, core implementation — all of it is in scope, all of it goes through a code review gate before landing. That gate is the whole design. We covered VaG two days ago making exactly this argument: an agent generating its own improvements is fine, an agent approving them is not.
The detail buried in the paper is the one worth remembering. A deployment called Hope has been running 161 days in free evolution across seven surfaces, under governed human communication. Not a benchmark run — a living agent that has been rewriting itself for five months. The authors are explicit that safety is the binding constraint here, since a self-modifying agent that can also pick which model API it calls is a different animal than one that can only edit its prompts. Roman Yampolskiy is on the author list, which tells you they know it.
Paper: https://arxiv.org/abs/2608.08311
Comments