OpenAI's coding agent — describe what you want built, no coding required.
The full script, word for word — 227 words, about 1 minutes to read. Current as at July 2026. AI tools change quickly; if something looks different when you try it, check the product's own help pages.
Codex is OpenAI’s coding agent. It can inspect a software project, explain it, change files and run checks. That makes it useful even when the first job is understanding the code, not writing more of it.
Codex is separate from Chat and Work. Its web client runs coding tasks in cloud environments; it is not inside normal ChatGPT web. In the desktop app, choose Codex from the top-left menu, then choose the project it may work with. Its history stays separate. Start by asking it to inspect and explain before changing anything.
Give it a bounded outcome: diagnose this error, add this validation, or write tests for this behaviour. Include the constraints and definition of done. Codex may need permission to run a command, edit files or reach a connected service. Approve only what you understand, and use the least access the task needs.
When it finishes, inspect the changed files, the explanation and the test results. Ask what it did not test and what could still fail. A green test is evidence, not proof that the change meets every requirement. Material work still needs human review before it is merged or deployed.
A good first task is low risk: explain an unfamiliar folder or improve one small test. Keep the scope clear, review every change and widen Codex’s responsibility only after it has earned your confidence.