Unrolling the Codex agent loop

Source: OpenAI Date: 2026-01-23 URL: https://openai.com/index/unrolling-the-codex-agent-loop

Summary

Title-only: A technical post from OpenAI explaining how the Codex agent loop works — the full cycle from receiving a task to completing code changes, including task decomposition, file reading, code writing, test execution, iteration on failures, and pull request creation. Published January 2026 as Codex was approaching its GA launch, this is both developer documentation and a transparency post about what autonomous AI coding agents actually do step by step.

Implications

The agent architecture transparency thread. “Unrolling” the agent loop is a deliberate transparency move — it demystifies what Codex does so that developers, security teams, and managers understand the autonomous actions an AI coding agent takes on their behalf. This matters for enterprise adoption: IT security teams need to understand what a coding agent can and cannot do before approving it for production codebase access.

Agent loop design as competitive differentiator. The specific design choices in Codex’s agent loop — how it handles file access, how it manages failures, how it creates pull requests — determine how safe and effective it is in practice. Publishing the loop design invites comparison with competing agent systems (Claude’s computer use, Devin, GitHub Copilot Workspace) and signals confidence in Codex’s architecture. Developers reading this post will evaluate whether Codex’s loop is more or less trustworthy than alternatives.