Ouroboros vs Claude Code: агент написал себя сам и достиг SOTA на Terminal-Bench 2.1

A researcher from AIRI published on August 3, 2026 a detailed breakdown on Habr of the self-evolving agent Ouroboros — a system built in pure Python that autonomously rewrote its own code over several months and achieved SOTA results on Terminal-Bench 2.1, CL-Bench, and OSWorld, matching Claude Code and Codex on the GAIA and SWE-Pro benchmarks.

How Ouroboros writes itself

Ouroboros is an agent loop in Python with direct access to git and the ability to overwrite its own runtime mid-execution. The closed loop works like this: the agent receives a task, executes it, analyzes the result, and makes edits to 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 hosted the first version on Google Colab rather than a local machine: the cloud environment's isolation protected against 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 the runtime, a safe rollback mechanism
• First tasks: drew cats, hung around on YouTube, changed desktop wallpapers
• Budget: autonomous token spending without constant developer oversight

What the benchmarks show

On three benchmarks — Terminal-Bench 2.1, CL-Bench, and OSWorld — Ouroboros reached SOTA-level results. All three assess different facets of agent capabilities: Terminal-Bench evaluates command-line work, OSWorld evaluates desktop control and GUI interfaces, and CL-Bench evaluates complex programming tasks. On GAIA and SWE-Pro, parity with Claude Code (Anthropic) and Codex (OpenAI) was recorded.

"I wouldn't have designed a harness like this myself, but autonomous

evolution and a pile of burned token money — sure," — the project's author, an AIRI researcher.

The current architecture of the system took shape not through manual design, but through dozens of autonomous iterations, each of which cost real tokens. Notably, Ouroboros achieved this without large computational resources — just a Google Colab cloud 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 — the source code remains unchanged throughout. Ouroboros works differently: the agent edits its own source code in a production cycle, and it is precisely the evolution of the architecture — not the cycling through prompts — that led to 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 remain closed or are not accompanied by open benchmark data. According to the author's Habr publication, the breakdown includes traces, figures for reproducibility, and "a few awkward stories from the audit" — cases where the agent made unexpected decisions during evolution.

The author reports that he now "does almost everything through it" — Ouroboros has transformed from a research project into a primary working tool.

What this means

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