OpenAI launches Codex: an AI agent that codes while you wait
Curated by the Inblix editorial team
OpenAI just took the next logical, if slightly terrifying, step: an AI that doesn’t just suggest code snippets but spins up an entire cloud sandbox, writes features, fixes bugs, and files pull requests all on its own. The new agent, called Codex, launched today as a research preview for ChatGPT Pro, Business, and Enterprise users, with Plus and Edu access rolling out soon. It’s powered by codex-1, a specialized version of the o3 model fine-tuned with reinforcement learning on real-world software engineering tasks.
The workflow is straightforward. You point Codex at a GitHub repository, type a prompt, and click a button labeled either ‘Code’ for generating new work or ‘Ask’ for questions about the codebase. The agent then boots up an isolated cloud environment, preloaded with your repo, and gets to work independently. OpenAI says a typical task takes between 1 and 30 minutes, during which you can watch its progress in real time. Once finished, it commits the changes and serves up terminal logs and test outputs as a verifiable paper trail. You can then review the patch, request revisions, or open a pull request directly.
OpenAI is pitching this as an agent that behaves more like a considerate human colleague and less like a bull in a china shop. The model was trained specifically to generate code that mirrors human style, adhere precisely to instructions, and keep running tests until it gets a passing result. To help it understand a project’s conventions, developers can drop an AGENTS.md file into their repository—a sort of README on steroids that spells out testing commands, codebase navigation, and standard practices. The company claims codex-1 already shows strong performance on internal benchmarks even without this custom scaffolding.
Security is the elephant in the room with any autonomous coding tool, and OpenAI appears to have taken it seriously. By default, the agent’s cloud sandbox has no internet access, limiting its interactions to the provided code and pre-installed dependencies. The model was also explicitly trained to identify and refuse requests for malicious software development, while still allowing legitimate low-level kernel work that might trip simpler filters. Still, the company is blunt about the limits, stating that manual review of all agent-generated code before integration remains essential. The approach mirrors an industry grappling with how to unleash powerful automation without handing over the keys to the kingdom.
💡 Key Takeaways
- Codex runs each task in a separate, isolated cloud sandbox with internet access disabled by default, limiting potential attack surfaces to the user's own repository and pre-configured dependencies.
- The underlying codex-1 model was trained with reinforcement learning to iteratively run tests until it achieves a passing result, mimicking a developer's own debugging loop.
- Developers can guide the agent's behavior through an AGENTS.md file, a plain-text instruction document that defines testing commands and coding conventions for a specific project.
- OpenAI implemented explicit training to reject malware development requests while preserving the ability to handle legitimate low-level engineering tasks that might otherwise trigger generic safety filters.
Keep reading: See related articles below for more coverage on this topic.
Get smarter about AI
The sharpest AI news, curated daily. Delivered free to your inbox.