Self-Evolving Agent Ouroboros Surpasses Codex and Claude Code on Major Benchmarks

The developer of the Ouroboros agent describes how they aimed to create a universal Jarvis-like agent but realized it was too difficult, so instead they gave an agent the ability to design and build itself in an autonomous evolution loop. Ouroboros is a Python-based agent loop with access to git, the right to rewrite its own runtime, safe restart on new versions, and rollback if something breaks. The first version was placed in Google Colab for safety, given a budget and autonomy. After several months, the agent, initially a simple desktop agent that drew cats and changed wallpapers, now achieves state-of-the-art results on Terminal-Bench 2.1, CL-Bench, and OSWorld, and parity with Claude Code and Codex on GAIA and SWE-Pro. The developer admits they could not have designed such a harness themselves, but autonomous evolution and token spending made it possible. The post includes benchmark numbers, reproducibility details, awkward stories from trace audits, and a brief overview of the architecture.