Previously published as Codex in Production by Jason Colapietro.

A coding agent can change files. Production work needs a system for deciding what it may change, where the work stays isolated, and what evidence counts as done.

Codex in Production is a practical operator's guide to running OpenAI's Codex across real repositories and multi-step engineering work. It covers the control layer around the agent: durable project instructions, isolated worktrees, reusable skills, bounded permissions, connector trust boundaries, long-running tasks, and verifiable completion.

Built for the harness, not one model. AGENTS.md, worktrees, approvals, and evidence gates do not change when the model does, so the control layer described here holds as the underlying model moves across generations, GPT-6 Astra included.

Fit check: this book assumes you already work in a terminal and a Git repository and have asked a coding agent for code before. It does not teach programming, Git, or prompting basics, and it is not a tour of screenshots.

Inside the guide

• Treat every task as a state machine with explicit transitions, evidence, and authority
• Choose the surface that matches the work: desktop app, IDE extension, CLI, or cloud
• Write the repository contract in AGENTS.md, with scope, hierarchy, and an adversarial instruction test
• Establish ground truth before the first edit with a ten-line orientation receipt
• Design the approval boundary: sandbox, approvals, and network access as separate controls
• Run one task per worktree with one owner, and retire worktrees without deleting evidence
• Delegate outcomes to subagents with a six-field contract instead of vague topics
• Turn repeated judgment into a skill, with routing tests and a versioned procedure
• Treat every connector, plugin, and MCP server as a trust boundary
• Keep long-running work legible with goal packets, checkpoints, and honest stops
• Review the claim, not just the diff, using an evidence ladder and an independent pass
• Automate the bounded middle with non-interactive runs, CI, the SDK, and App Server
• Recover from an incident without making it larger
• Install the whole operating system in thirty days, week by week

How it teaches

Every chapter follows one fictional but reproducible repository, Relay, from first inspection to a production-ready change. Each chapter states a real operator problem, the Codex mechanism that addresses it, a workflow you can rerun, a failure mode with its recovery path, and a drill or acceptance gate. Product facts cite current official documentation with a research date. Observed behavior is recorded in a ledger and kept separate from patterns that are labeled as exercises.

What you take away

Five appendices hold the working templates: repository contracts for solo, monorepo, and team settings; operator packets for orientation, delegation, long-run goals, connector threat models, verification matrices, and incident handoffs; a release-state vocabulary; the version and evidence ledger; and the thirty-day installation checklist.

Who it is for

Developers, technical founders, engineering leads, and hands-on operators who need agent work they can direct, inspect, and verify. The aim is not to remove judgment from software work but to make agent work easier to bound and prove.

From the author of The Claude Code Bible and the open-source Suede Creator Skills pack. Both books belong to The Operator's Library and can be read in either order.

An independent guide. Not affiliated with or endorsed by OpenAI.