The Openai agent announced on May 16 is capable of code, correct and document in total autonomy using a Github deposit.
Openai continues to hit the code strongly. The San Francisco start-up presents this Friday, May 16 Codex, a new autonomous code agent. Operating in an isolated cloud environment, Codex can not only write code, but also answer questions on a code base, correct bugs and even offer requests.
An autonomous code agent
Powered by Codex-1, an optimized version of the O3 model specially designed for software engineering, Codex is driven by apprenticeship by strengthening on actual programming tasks. Each Codex agent runs in a dedicated cloud container, preloaded with the user code deposit and completely disconnected from the Internet. The agent can read and modify files, run tests. The tasks are carried out within 1 to 30 minutes depending on their complexity.
Codex can generate new features in the code, answer questions about the architecture of a project, carry out complex refactoring, write unit tests or even technical documentation. The agent is also able to go beyond the simple generation of code: he analyzes recurring problems in the code, configuration inconsistencies, and offers improvements. It can detect and correct bugs, suggest performance optimizations, and even generate requests for requests.
The agent integrated directly into Chatgpt
Codex fits directly into the chatgpt interface via a new dedicated lateral section. The developers can interact with the agent in two ways: by clicking on Code to launch a new development task, or on Ask to ask questions about an existing deposit. Each task is performed independently in an isolated cloud environment with the project code. The user can follow the progress in real time and receive a report of the operations carried out. Once the task is completed, it is possible to request revisions, open a request directly on Github, or integrate changes into the local environment. The agent can also be guided by an agent file.MD (on the principle of README) by specifying code conventions, test instructions and project practices.
The agent is launched from this 16 but for users Chatgpt pro, company and teams, with an extension planned for more and Education in the coming weeks. The service will initially be offered for an additional cost but with request limits. It will also be possible to buy additional requests.
The agent will also be available in the API under the codex-mini-latest name. The model is accessible via the API Responses and is billed at $ 1.50 for 1 million entry tokens and $ 6 for 1 million output tokens, with a 75%cache discount. It is possible to use it for simple code factoring tasks, documentation generation or even bug correction.
For Openai, the developer profession is transformed
OPENAI imagines a future (close, very close?) Where would the developers strategically pilot their projects while delegating repetitive or technical tasks to intelligent agents. The objective is to allow engineers to focus on innovation and creativity, letting AI manage the more routine aspects of software development. Ultimately, the start-up provides for a fluid collaboration between humans and agents, where they can work in asynchronous and offer proactive updates.
Today, the agent saves time and automate certain tasks, but he does not replace developers. OPENAI remains voluntarily vague on the long -term impact but ends his laius with an eloquent thought: “It’s just The Beginning.”
