Ouroboros vs Claude Code: An Agent That Wrote Itself Hits SOTA on Terminal-Bench 2.1

On August 3, 2026, an AIRI researcher published a detailed write-up on Habr of Ouroboros, a self-evolving agent written in pure Python. The system spent several months autonomously rewriting its own code and reached SOTA results on Terminal-Bench 2.1, CL-Bench and OSWorld. It also matched Claude Code and Codex on the GAIA and SWE-Pro benchmarks.

How does Ouroboros write itself?

Ouroboros is an agentic loop in Python with direct access to git and the right to overwrite its own runtime while it works. The closed cycle looks like this: the agent receives a task, executes it, analyzes the result and edits its own source code, then safely restarts on the updated version. If a new iteration breaks something, the system automatically rolls back to the previous one.

The author deliberately housed the first version in Google Colab rather than on a local machine: the isolation of the cloud environment protected against the unpredictable consequences of self-evolution. The system started modestly.

• Language: pure Python with no third-party agent frameworks
• Environment: Google Colab, isolated from the developer's local machine
• Tools: git, the right to overwrite its runtime, a safe rollback mechanism
• First tasks: drawing cats, getting stuck on YouTube, changing the desktop wallpaper
• Budget: autonomous token spending without constant oversight from the developer

What the benchmarks show

Ouroboros reached SOTA-level results on three benchmarks: Terminal-Bench 2.1, CL-Bench and OSWorld. Each measures a different facet of agentic capability: Terminal-Bench tests command-line work, OSWorld tests control of the desktop and GUI interfaces, and CL-Bench tests complex programming tasks. On GAIA and SWE-Pro, the agent shows parity with Claude Code (Anthropic) and Codex (OpenAI).

"I would never have come up with a harness like this myself, but

autonomous evolution and a pile of money burned on tokens did," says the project's author, an AIRI researcher.

The system's current architecture did not emerge from manual design. It took shape over dozens of autonomous iterations, each of which cost real tokens. Notably, Ouroboros got here without major compute resources: only the Google Colab environment and a budget for API requests.

How this approach differs from ordinary agents

Most agent systems improve through prompt engineering or fine-tuning on new data, while their source code stays unchanged. Ouroboros works differently: the agent edits its own source code in a production loop, and it was this evolution of the architecture, not prompt trial and error, that led to the SOTA results.

This makes Ouroboros one of the few publicly documented examples of genuine agent self-modification with verifiable numerical results. Most similar experiments either stay closed or ship without open benchmark data. According to the author's Habr post, the write-up includes traces, figures for reproduction and "a few awkward stories from the audit," cases in which the agent made unexpected decisions during its evolution.

The author reports that he now does "almost everything through it": Ouroboros has gone from a research project to his main working tool.

What this means

Ouroboros is practical evidence that agent self-evolution can deliver competitive results without multimillion-dollar compute budgets. A single researcher with a Colab environment reached SOTA in an arena where labs with teams and server infrastructure compete. The next question is how to manage such evolution at scale and how to prevent unwanted behavior from an agent that writes itself.