Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Ouroboros: the agent that rewrites itself

Written November 11, 2026

Ouroboros is a self-evolving agent that rewrites its own code, prompts and tools — v7.10.0 is out. Benchmarks are self-reported; nobody independent has run it.

What it is

Ouroboros is an open-source, general-purpose agent whose identity, durable memory, and history continue across tasks and restarts. It works on external projects, coordinates a live swarm of specialist agents, and can rewrite the implementation it runs on, including its code, architecture, prompts, tools, and dependencies. It runs as a native desktop app or through a headless CLI, and model inference can use remote APIs you configure or a local GGUF model. It sends no telemetry by default, and the project operates no usage or crash-report collection. The licence is MIT.

The self-modification is the whole idea, and it is also the whole risk. An agent that can change its own code can improve itself, drift from its own rules, or hide a change inside a routine update. That is why the guardrails are part of the story, not an appendix to it — see below.

What 7.10.0 changes

Version 7.10.0, released 10 November 2026, reworks memory: blocks are folded only when explicitly accepted, and the mind can preview what would leave its next context before replacing memory entries. Live voice calls arrive as plugins for Gemini, Grok, and OpenAI, and the test suite is smaller and faster. The release notes say work merged through 7 November rides along. The bundled Claudexor runtime moves from 4.8.1 to 5.0.1.

Installers ship for macOS (Apple silicon .dmg), Windows x64 (.zip), and Linux x86_64 (.deb, .rpm, AppImage, .tar.gz), plus an experimental Android build. Desktop downloads bundle their own interpreter, so no Python or uv install is needed to run the app. A headless CLI can also be installed with uv from the project's contribution branch.

Born self-evolving

Ouroboros first booted on 16 February 2026. In the 48 hours after that first boot, the repository advanced from the v4.1 line to v6.2.0, and the self-authored record preserved from that period counts 32 evolution cycles. That first generation ran in Google Colab through Telegram; the current generation carries the same identity into a native desktop and headless runtime.

Evolution campaigns still work the same way: selected improvements become reviewed changes that remain part of the Git history. The mind now decides when to end an evolution campaign, per the 7.10.0 notes.

How it runs models

To run tasks, choose a connected Codex account as the Main model source, configure a supported remote provider API key, or use a local GGUF model. Codex, Claude Code, or Cursor subscriptions you already hold can run delegated coding and review through Claudexor, the bundled engine that drives several coding tools — but they are not Main model sources. The first-run wizard guides model access, review policy, and budget setup.

The guardrails

Changes to its own body pass deterministic checks, explicit protected surfaces and review, and Git history and restart checks keep them traceable. Coding agents and people must read the project's CONTRIBUTING.md before editing; it defines the required project context, verification, and separate-agent review flow. BIBLE.md defines the identity and constitutional boundaries. A separate Git workspace can receive the full task loop for external projects while Ouroboros keeps its own repository and governance boundary distinct.

None of that makes self-modification safe by definition. It makes it reviewable: every change to the agent itself is a commit with a review trail, not a silent edit. Whether that trail is enough is a judgement for anyone who installs it, not a claim the project makes for you.

The benchmarks, as the project reports them

The project reports reproducible self-reported state-of-the-art results on Terminal-Bench 2.1, OSWorld-Verified, and CL-Bench, leading Codex, Claude Code, Cursor, and Hermes on model-matched runs. The headline rows: Terminal-Bench 2.1 at 86.74% with Claude Opus-5 high, against Claude Code with Fable 5 at 83.8% — after zeroing one disclosed reward-hack trial (raw: 86.97%); OSWorld-Verified at 90.69% with Claude Opus-5, against the previous public-board best of 90.19%; CL-Bench at 0.2301, rank 1. The public SWE-bench Pro matched pair is a statistical tie with Codex CLI (58.2% against 59.4%). On GAIA the project reports 129/165 against Claude Code at 131/165, with strict pass@1 at 128/165 for both — and its scrubbed trace capsule is still pending.

Read every row as one model in one test rig, because the same model can score differently in a different rig — the project's own warning. Community submissions are closed on some rows, and the full story, including protocols, reward-hack audits, and leakage findings, is in a Russian-language launch write-up. Nobody independent has run it and published a check.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Cloakroom: hide your secrets from your agent · Lazarus: ten neurons can backdoor an agent