# Ouroboros

An Agent OS that turns non-deterministic coding agent work into a replayable, observable, policy-bound process through interview, execution, and evaluation stages.

Ouroboros runs a structured, specification-first workflow instead of ad hoc prompting: it interviews the user about a vague request, crystallizes it into a testable specification, executes against that specification, evaluates the result, and evolves from there, reporting an ambiguity score as it narrows down what is being asked for. It is designed to work the same way across many different coding agent runtimes and hosts, including Claude Code, Codex CLI, OpenCode, Hermes, Gemini, Kiro, and Copilot, keeping the interview and evaluation engine constant even though each host and task differs.

It installs with a single shell or PowerShell command that also installs Git and uv where needed, after which a one-time `ooo setup` step runs inside the chosen coding agent. The project is organized as three separate repositories forming a stable OS layer, an application layer of domain workflows, and the CLI shell a user interacts with directly.

It is written in Python and distributed under the MIT license.

Website: https://ouroboros.page/