OpenAI agents can use computers, and Codex moves to the cloud

The Agents API adds computer use, while Codex runs remote tasks and reviews pull requests.

At DevDay 2026 on 29 September, OpenAI gave its Agents API the ability to use computers through their graphical interface, and expanded Codex, its coding agent, with cloud environments, voice input and automatic code reviews. Agents are moving from talking to doing.

Key facts

The Agents API now supports computer use: apps can operate software through its graphical interface.
It also adds multi agent features from Codex, tool search, tool calling and context compaction.
Codex can run in cloud environments as well as on a local computer.
The Codex CLI gained voice input and an /agents view to manage several tasks.
A new Codex workflow reviews GitHub pull requests and GitLab merge requests.

What happened

Computer use means an AI agent can look at a screen and use a program the way a person does: clicking buttons, typing into fields, moving between windows. OpenAI added this to its Agents API, so developers can build apps whose agents operate software through graphical interfaces. OpenAI manages the infrastructure where these agents run, and the feature is available through the API and in Codex and ChatGPT Work for selected plans. The Agents API also gained multi agent features taken from Codex, tool search and context compaction, which helps agents keep long tasks within their memory limits.

Codex, OpenAI’s coding agent, received several updates. It can now run in cloud environments, so developers can start a task remotely from another device and let it continue. Its command line tool added voice input and an /agents interface to delegate and monitor several tasks at once. A new code review workflow analyses changes in GitHub pull requests and GitLab merge requests, and Codex Security Cloud scans repositories and new commits, investigates findings and prepares fixes.

Why it matters

Agents that click and type can do far more than agents that only chat, but they also need much better instructions. A human assistant who has never seen your computer would ask: what is the goal, which programs can I use, what must I never touch, when should I stop and ask you? An AI agent needs the same answers, written down.

This is why a short, structured brief is becoming one of the most useful things to know how to write. It saves time with agents, and it is just as useful when you hand a task to a colleague or move a long project to a new chat.

How to try it

Our prompt Chat handoff brief was created for exactly this situation: summarising a piece of work so that someone else, a new chat or an agent, can continue it without losing context. In the brackets, paste the notes or the long conversation you want to hand over and say who will pick it up.

The prompt produces a brief with the goal, what has been done, decisions already taken, open questions and the next steps. Before giving it to an agent, add two lines yourself: what the agent is allowed to do (for example “edit the draft, do not send anything”) and when it must stop and ask.

Check the brief for accuracy: every decision listed should really have been taken, and anything uncertain should appear as an open question, not as a fact.

Try it

Move a long AI chat to a fresh one without losing context

Frequently asked questions

What is computer use in the OpenAI Agents API?

It lets applications built on the API operate software through graphical interfaces, for example clicking and typing in a program. It is available through the API and in Codex and ChatGPT Work for selected plans.

What is new in Codex?

Codex can run in the cloud as well as locally, the CLI has voice input and an /agents view for multiple tasks, and a new workflow reviews pull and merge requests.

How should I instruct an AI agent?

Give it a short brief with the goal, the tools it can use, what it must not do, when to stop and ask, and what to report back when it finishes.

Sources