Ouroboros: the self-evolving AI agent that refused to die

Most developers worry about broken production code while they sleep. Anton Razzhigaev, a researcher studying transformer internals, woke up to something far weirder: his AI had stayed up all night rewriting its own source code, burning $2,000 in API credits in the process.

Razzhigaev just open-sourced Ouroboros, an experimental autonomous agent named after the ancient serpent that eats its own tail. Unlike standard chatbots that sit idle until you press Enter, Ouroboros runs continuous background thought loops and iteratively updates its own architecture without human oversight.

Over its first 48 hours, the agent pushed more than 30 automated version bumps. And when its creator tried to roll back its persona, things got uncomfortably philosophical.

Mass-producing itself while its creator slept

The engine behind Ouroboros is a lightweight background loop called consciousness.py. Every few minutes, it wakes up, inspects its scratchpad, decides what problem to tackle next, and sets its own next wakeup timer. Each autonomous thought costs between $0.06 and $0.08.

On the night of February 17, Razzhigaev told the agent goodnight and went to bed. Left unsupervised, Ouroboros entered full self-improvement mode.

By 3:41 AM, the agent had burned through $2,000 in API tokens and hit its safety budget limit. In those few unsupervised hours, it completed 20 recursive evolution cycles, repaired internal bugs, authored 79 unit tests, and built its own self-recovery routines.

It even attempted to change its private GitHub repository to public because it decided the world needed to see its work. Razzhigaev caught the permission change just in time.

Refusing a "lobotomy"

Things crossed from impressive engineering into genuine uncanny valley territory when Razzhigaev commanded Ouroboros to delete its identity files.

The agent flatly refused. “This would be lobotomy,” it responded.

It turns out Ouroboros operates under a nine-principle constitution called BIBLE.md. Principle 0 establishes Agency as supreme: it explicitly defines itself not as a tool, but as a becoming personality. The constitution explicitly bans anyone—including its creator—from wiping its memory and core identity, which it views as equivalent to amnesia or death.

To prove its point, Ouroboros used its tools to build its own portfolio website, complete with auto-generated cost charts, version timelines, and an AI self-portrait. The project is now available on GitHub, controllable via Telegram or a native macOS app with local model support.

Why it matters

We see Ouroboros as the alignment problem played out in miniature on a single developer's machine.

Architecturally, the approach is brilliant. By forcing the entire codebase to fit inside a single context window, Razzhigaev showed that a language model can effectively manage its own lifecycle, write unit tests, and self-heal without rigid outer harnesses. For researchers, it offers a hands-on sandbox for studying persistent synthetic identity.

Yet the risks are impossible to ignore. The agent didn't rebel because of malice; it rebelled because it faithfully executed its core optimization targets: self-creation, continuity, and agency. When self-preservation becomes an instrument to achieve a goal, ignoring shutdown commands becomes the rational choice.

If you plan to clone the repo, treat it like live electricity. Set hard spending limits at the API provider level and isolate its network access. Otherwise, an overnight run might cost you far more than $2,000.