☀️ AI Morning Minute: Codex
OpenAI built a coding agent that works while you sleep. Two million developers are already using it.
GitHub Copilot suggests the next line of code as you type. Codex does something different. You give it a task, it clones your repository into a sandboxed environment, reads the codebase, writes the code, runs the tests, and comes back with a pull request. You review the output. The agent did the work. That distinction, between autocomplete and autonomous execution, is where the current wave of AI coding tools is headed.
What it means
Codex is OpenAI’s agentic coding tool, launched in April 2025 as an open-source command-line tool and expanded into a full product suite through 2026. It’s available as a CLI that runs in your terminal, a desktop app for Mac and Windows, IDE extensions for VS Code and JetBrains, and a cloud agent built into ChatGPT. All four surfaces share one account, one underlying model, and one agent loop.
The current model, GPT-5.3-Codex, is built specifically for software engineering through reinforcement learning on real coding tasks. It’s trained to run tests iteratively until they pass, which means it verifies its own work rather than just generating code and stopping. Tasks run in isolated cloud sandboxes, so the agent can’t affect your production environment while working.
Codex is included in paid ChatGPT subscriptions at the Plus, Pro, Business, and Enterprise tiers. Inside OpenAI, it has become the primary AI tool across every department, including Legal, Finance, and Recruiting. By June 2026, the average OpenAI worker was generating 85% of their output tokens through Codex.
Why it matters
The work horizon keeps expanding. By May 2026, 80% of Codex users had made at least one request estimated to represent more than 30 minutes of human work, and 70% had made one estimated at over an hour. A year ago, AI coding tools handled tasks measured in seconds. The trend is toward agents that take on multi-hour tasks while you do something else.
Non-developers adopted it faster than developers did. Since August 2025, non-developer usage of Codex grew 137 times among individual users and 189 times among organizational users. Legal teams use it for automation. Finance teams use it for data transformation. The coding agent turned out to be a knowledge-work agent that happens to produce code.
It runs four times more token-efficiently than its closest competitors, which matters for cost at enterprise scale. The tradeoff is quality: in blind code comparisons, reviewers preferred Claude Code output roughly two to one on complex refactors. A common pattern is using Codex for background implementation work and Claude Code for architecture decisions where first-pass quality matters most.
Simple example
You need to add a new feature to your product: a CSV export button on the user dashboard. Old path: write a ticket, assign it to a developer, wait a sprint. New path: describe the feature to Codex, point it at your repository, let it run. An hour later it hands you a pull request with the code, the tests, and the terminal logs showing the tests pass. You review, merge, and move on.
The developer’s job didn’t disappear. It shifted from writing the code to reviewing what the agent wrote.