How to use Codex to build an agent-written product
The team reports a shipped internal beta with daily internal users and external alpha testers, roughly 1,500 merged pull requests and about a million lines of agent-written material.
openai.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Build [product] in [repository] from [templates]. Scaffold CI, formatting, package setup and agent instructions, then break [goals] into implementable tasks. Open pull requests, review changes locally, obtain additional agent reviews, and iterate on feedback until reviewers are satisfied. Report blockers requiring human direction.
Reach people exploring AI agents
See placements →What happened
The team reports a shipped internal beta with daily internal users and external alpha testers, roughly 1,500 merged pull requests and about a million lines of agent-written material.
Reported by i0exception on BlogHow the agent did it
- Generate the initial repository scaffold from an empty Git repository.
- Break larger product goals into smaller building blocks.
- Prompt Codex to implement tasks and open pull requests.
Where a person stepped in
Humans specified tasks and built the environment and feedback loops. Human pull-request review was possible but not required.
A person set the goal and reviewed the result.
OpenAI describes a five-month experiment in building and shipping an internal software beta without manually written code. Codex generated application logic, tests, CI configuration, documentation, observability, and internal tooling. The article reports internal daily users and external alpha testers, so the account concerns a used product rather than a code-generation demonstration.
The team began with an empty repository. Codex CLI, using GPT-5 and existing templates, generated the repository structure, CI setup, formatting rules, package-manager setup, and application framework. Codex also wrote the initial AGENTS.md. Engineers then broke larger goals into building blocks and prompted the agent to construct the capabilities needed for subsequent tasks.
For individual changes, an engineer described a task and allowed Codex to open a pull request. The team instructed it to review changes locally, request additional local and cloud agent reviews, respond to feedback, and iterate until reviewers were satisfied. It gathered context through gh, local scripts, and repository skills. Human PR review was possible but not required.
The source reports roughly 1,500 merged pull requests and about a million lines across code and supporting material. These are the team's reported totals, not independent verification. Who it's for: engineering teams prepared to define intent, repository structure, and review feedback loops for an agent-led development process.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · i0exception
Frequently asked questions
How did the repository get started?
Codex CLI using GPT-5 generated the initial scaffold from an empty repository with guidance from existing templates, including CI, formatting, package setup and the initial AGENTS.md.
What review process did changes follow?
Codex was instructed to review locally, request additional local and cloud agent reviews, respond to feedback and iterate until agent reviewers were satisfied. Human PR review was allowed but not required.
More details
Remaining steps
- Request local and cloud agent reviews of each change.
- Iterate on feedback until agent reviewers are satisfied.
About this account
Codex wrote the product and supporting engineering files, opened pull requests and participated in repeated agent-to-agent reviews.
Platform: Blog. Attribution: https://openai.com/index/harness-engineering/. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: i0exception on Blog
Source published 12 Feb 2026
Source checked 14 Oct 2026
Tools and topics
Codex CLI · GPT-5 · gh · Repository skills
Agent-written code · Pull requests · Repository setup
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.