Build Anything with Codex, Here’s How

Codex is positioned as a production-grade coding agent that can chew through real GitHub issues in minutes—then open pull requests with targeted diffs—while running tasks asynchronously in the background. The practical payoff shown here is a workflow where dozens of issues that might take days for a human team member can be processed in parallel, with developers reviewing and merging only the changes that pass checks.

Access starts at chatgpt.com/codex, with the creator recommending the Team plan as a lower-cost path than a $200/month option. Codex runs on a model called Codex 1, described as OpenAI’s most capable coding model so far, built by fine-tuning a prior system (referenced as “03”) on senior-level production coding practices. The reinforcement process is framed as mirroring real engineering work: writing unit tests, adding comments, splitting changes into smaller files, and understanding the codebase structure.

Setup centers on connecting a GitHub repository and choosing a branch. The workflow emphasizes creating “environments” for different codebases (the example includes a testing setup and a production-level codebase used by over 50,000 users). A key safety rule is repeated: Codex should not run on a production branch. Instead, changes go to staging (dev) or a personal branch for experiments.

A demonstration uses a real GitHub issue from the Vectal codebase: persisting the last selected AI model and chat agent mode in browser local storage. The transcript stresses that vague ideas aren’t enough—Codex prompts need to specify what to change, what to avoid, and how to implement it “in the simplest and cleanest way possible,” ideally with minimal line changes. After submitting the task via the “code” option, Codex completes it quickly (reported at 2 minutes 50 seconds), produces a concise diff (adding 21 lines and removing 7 in a single file), and pushes a new pull request. Initial lint failures are traced to missing advanced environment dependencies, which leads into the next phase: configuring environment execution.

Advanced environment setup is treated as the difference between smooth autonomous runs and repeated failures caught by deployment checks. The creator walks through editing an environment to install backend and frontend dependencies using repo-appropriate terminal commands (e.g., pip install -r requirements.txt for the backend and npm install for the frontend). The setup also includes adding an OpenRouter API key so Codex can run tests that depend on external model access. When a task fails due to incorrect dependency installation context (a “CD back to root” mistake), the fix is iterated by updating the environment commands and re-running tasks on a non-production branch.

The final quality-of-life goal—restoring the saved model/mode after merging—requires debugging when the preference persistence doesn’t initially work after deployment. A follow-up Codex run investigates why the app still defaults to the original model and agent mode, then later succeeds. After the corrected pull request is merged, the transcript shows the preference sticking across reloads (switching to Gemini 2.5 Pro and chat mode persists after closing and reopening).

Overall, the transcript sells Codex as an agentic engineering tool: launch many tasks at once, let it run installs and tests, review diffs and PRs, and merge when checks pass—turning routine engineering work into a parallelizable background process rather than a sequential, human-driven grind.