Ouroboros outperforms Codex, Claude Code, Cursor and Hermes on three terminal-agent benchmarks

The Ouroboros agent, a pure-Python automation loop, placed first on Terminal Bench, OSWorld and CL-bench, the three widely used benchmarks for measuring a model’s ability to operate in a terminal, use a computer and write code. The results were released in full together with the execution scripts, evaluation scripts and complete trace logs, all under an MIT license that permits commercial use, modification and redistribution.

The author states that all of his development and research work now proceeds through Ouroboros, describing the tool as a rare example of a self-developing agent. He reports that running the full suite of benchmarks cost an amount equal to buying an apartment, illustrating how expensive systematic evaluation of agents at this level can be.

The source code is hosted on GitHub, with documentation on the Habr channel and official installers. Unlike many models that publish only weights, Ouroboros provides complete source code and execution data that enable full reproducibility without reliance on external cloud services. The permissive license and absence of proprietary components make Ouroboros suitable for local execution on relatively modest hardware, provided sufficient memory is available for the runtime context. The author invites bug reports and pull requests, promising that all contributions will be reviewed through Ouroboros itself.