OpenAI Codex: from autocomplete to agent

Codex is OpenAI's stake in the agent-coding war — and it works in two flavors: Codex web, a cloud agent you dispatch whole tasks to in parallel (spec → branch → PR), and Codex CLI, an open-source Rust terminal agent. Both run on the GPT-5 lineage models, both are included with ChatGPT plans — OpenAI betting that the agent is the product now.

What it actually does

• Codex web — describe a task, Codex works in a cloud container: edits, tests, commits a PR for your review. Multiple tasks in parallel, each isolated.
• Codex CLI — open-source terminal agent (npm i -g @openai/codex) with approval modes; works on your local repo, asks before destructive actions.
• GPT-5 models — reasoning modes designed for long agentic chains; strong on tool use and multi-step code work.
• Plan tiers — included with ChatGPT Plus/Pro/Enterprise (limits vary); API available for teams.

How it actually fits

• Web is for delegating: 'refund the tax logic to spec X' — you come back to a PR.
• CLI is for pairing: interactive, inspect the diff, approve the commands, steer it live.
• Codex shines when the task has a clean definition of done — tests it can run, requirements it can check — agent autonomy lives and dies on that.

Who it's for

• ChatGPT subscribers who want agent help without another subscription.
• Teams who want parallel task dispatch (multiple bugs/features at once).
• Anyone curious how OpenAI's agent stack actually behaves on real code.

What to check

• Treat web-Codex like a diligent junior: review every PR like you own it — because you do.
• The CLI is open source — audit it, and check parity with the web agent before you pick a lane.
• Limits on the free tiers are real — read the current plan details; they change often.

Codex made 'agent' the default interaction pattern for OpenAI — not chat, not completion: delegation. Repository: github.com/openai/codex.