An OS that interviews you before letting AI agents write code

What it does

Ouroboros is a local-first runtime layer—self-described as an “Agent OS”—that wraps AI coding agents in a policy-bound execution contract. It replaces ad-hoc prompting with a structured loop: a Socratic interview exposes hidden assumptions, crystallizes them into an immutable Seed spec, executes via Double Diamond decomposition, and verifies output through a three-stage evaluation gate. Every action is ledger-recorded and replayable across Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Hermes, Gemini, and Kiro CLI.

The interesting bit

The architecture literally eats its own tail: evaluation output feeds back as input for the next generation’s Seed specification until ontology similarity converges at 0.95. It treats intent—rather than prompts—as the core infrastructure primitive, attempting to turn non-deterministic agent work into a replayable execution contract.

Key highlights

• Multi-runtime adapter for Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Hermes, Gemini, and Kiro CLI
• Three-layer stack: ouroboros core (kernel), ouroboros-plugins (domain workflows), and ourocode (terminal TUI shell)
• Immutable Seed specs with acceptance criteria; the README cites ambiguity scores (e.g., 0.15) and a gate that blocks execution if intent is too unclear
• Three-stage evaluation: Mechanical (free), Semantic, and Multi-Model Consensus
• MCP server integration and ledger-based provenance for auditable, scoped plugin permissions

Verdict

Developers who view AI coding as reproducible engineering rather than improvisational chat will find the spec-first model compelling. If you prefer quick, ad-hoc prompting, this level of structure will feel like overkill.