OpenAI releases Codex, an AI coding agent that can run multiple tasks in parallel on a cloud-based platform
OpenAI has released a research preview of Codex , a cloud-based software engineering agent that can perform multiple tasks in parallel, such as developing features, answering questions about a codebase, fixing bugs, and suggesting pull requests. Codex is available for ChatGPT Pro, Enterprise, and Team users, with Plus and Edu coming soon.
Introducing Codex | OpenAI
ChatGPT is getting an AI coding agent | The Verge
https://www.theverge.com/command-line-newsletter/668251/chatgpt-is-getting-an-ai-coding-agent
ChatGPT rolls out Codex, an AI tool for software programming
https://www.bleepingcomputer.com/news/artificial-intelligence/chatgpt-rolls-out-codex-an-ai-tool-for-software-programming/
Codex has been trained on real coding tasks in a variety of environments using reinforcement learning, and OpenAI claims that it can generate code that closely resembles human style and pull request preferences. Codex also has the ability to follow instructions precisely and run tests iteratively until it gets a passing result.
Users can access Codex from the ChatGPT sidebar and assign coding tasks by entering a prompt and clicking the “Code” button, or use the “Ask” button if they have a question about the codebase.
A Codex is configured via an AGENTS.md file, a text file similar to a README.md file placed in the repository, that tells the Codex how to navigate the codebase, commands for testing, standard practices for the project, etc.
Codex's core feature is its parallel processing ability, which allows it to process multiple tasks simultaneously. Each task runs in an isolated environment in the cloud, preloaded with the user's repository. As a technical foundation, it is equipped with 'codex-1', an optimized version of OpenAI o3 for software engineering.
Codex can read and edit files, run commands like test harnesses, linters, type checkers, etc. Each task is processed independently in an isolated environment, and the codebase is pre-loaded.
Tasks take anywhere from 1 to 30 minutes to complete, depending on their complexity, and users can see the progress in real time. After a task is complete, Codex commits the changes in the environment and provides proof of each step of the process through terminal logs and test output. Users can also review the results, request additional fixes, create a GitHub pull request, or integrate the changes into their local environment.
Codex runs in the cloud in a secure, isolated container with internet access disabled during task execution, and real-time pairing with agents and task delegation is limited to code explicitly provided through the GitHub repository and pre-installed dependencies configured by the user in a setup script.
In internal benchmarks and coding evaluations, OpenAI claims that codex-1 has shown high performance even without an AGENTS.md file or custom scaffolding. In
Codex is also designed with a focus on security and transparency, with built-in functionality to validate output. Quotes, terminal logs, and test results allow users to review their work, and in the event of uncertainty or test failures, Codex explicitly communicates these issues to help users make informed decisions about how to proceed. It is recommended that all agent-generated code be manually reviewed and validated before integration or execution.
OpenAI said, 'Our future plans are to introduce more interactive and flexible agent workflows,' aiming to enable developers to provide guidance during tasks, collaborate on implementation strategies, and receive proactive progress updates. In addition, at the time of writing, Codex can connect to GitHub, but in the future, OpenAI has revealed that it aims to be able to assign tasks from Codex CLI, ChatGPT Desktop, and even external tools.
Related Posts:
in Software, Web Service, Posted by log1i_yk