Codex keeps coding in the cloud after your laptop closes

Codex can now keep working on a task after you close the laptop, let you check on it from your phone, and start every new task from a setup your team has already approved, TechCrunch reports from OpenAI's DevDay on Tuesday. Until now, each cloud task ran in a single-use sandbox. With this update, those environments become persistent workspaces you can configure and reuse.

At a glance

• OpenAI used DevDay to announce several Codex updates at once: reusable cloud environments, a refreshed CLI with voice control, code review in the ChatGPT desktop app, and Codex Security Cloud.
• Codex inspects your repositories and prepares dependencies and tools, then you publish that setup once. Each later task starts from it in a separate workspace that keeps running while your computer sleeps.
• Neither source gives prices, usage limits or rollout dates for the new environments, and the Decisions API is described only as using Luna, with no detail on what Luna is.

If you have not been following Codex, it is OpenAI's software engineering agent, and it could already spin up cloud tasks to do work remotely. Until this update, each of those tasks ran as an isolated remote sandbox. The change is about persistence and configuration: the setup now carries over from one task to the next.

Codex prepares a project once, and every task starts from that published environment

You start by choosing repositories. Codex inspects the project, works through setup, and asks you for any missing details or access. It prepares dependencies and tools, and you review the checks before you publish the environment. After that, tasks start from the published environment, each in its own workspace, and you update the setup as the project changes.

OpenAI says the reusable environments are meant to make tasks start faster and to give teams a shared workspace with approved settings and permissions. You can run Codex on your computer, remotely from a phone, or in the cloud, and review results from the web, mobile or desktop app. When a task finishes, you inspect the changed files, ask for follow-up changes, and commit or open a pull request.

The Codex CLI takes voice commands, and code review moves into the ChatGPT desktop app

The command-line tool got its own refresh. You can now start and direct tasks by voice, and a new /agents view is meant to make it easier to delegate work and track several tasks at once. From the terminal you can also start a cloud task, list recent cloud chats and check their status.

Smaller changes cover everyday work: editing prompts, resuming sessions and using Git worktrees, which let a single repository keep several branches checked out side by side. OpenAI is also rolling out what it calls a cleaner terminal interface, designed to keep longer sessions readable.

Codex can also be pulled into a new code review experience in the ChatGPT desktop app. There you read summaries, explore code changes and ask Codex about potential issues before you leave feedback on a GitHub pull request or a GitLab merge request.

Automatic code reviews go a step further: Codex does an initial review while you are away from your computer, so a first pass is ready before you open the change yourself.

Codex Security Cloud scans GitHub repositories on demand, on a schedule and on every new commit

On the security side, OpenAI is rolling out Codex Security Cloud, a set of tools for hardening infrastructure. It scans entire GitHub repositories on demand, on a regular schedule, or whenever new commits arrive. Codex then investigates the findings, removes duplicates and prepares fixes in the cloud, even when you are not active and your laptop is closed.

The toolset also gives access to models from Daybreak Blue, OpenAI's cybersecurity initiative, and you don't have to submit a separate application to use them. On the API side, OpenAI announced a Decisions API for real-time decision-making, which uses Luna to answer user-defined questions with predefined answers. An updated Agents API adds computer use and can use Amazon's Bedrock Managed Agents to build OpenAI agents that run entirely on AWS.

One published environment feeds many tasks, each with its own working state

The design keeps two things apart: the setup and the work. The setup is the repositories, dependencies, tools and network rules you publish once. The work is each task's own working state, kept separate, so two tasks started from the same environment do not trip over each other's files. Think of a restaurant kitchen where the ingredients are prepped once in the morning and every cook then works at their own station.

Access to outside services goes through the environment settings. You give Codex the package registries, APIs and private services the project needs, configure network access and supply credentials there. For credentials, Codex Cloud uses network secrets for allowed HTTPS services, and you can limit which network destinations the environment can reach at all.

The sources don't cover the practical questions. Neither gives prices, usage limits or rollout dates for the reusable environments, and the Decisions API names Luna without saying anything more about it. In our view, pairing network secrets with a restricted list of destinations is the right default for an agent that, by design, keeps running while your computer is asleep and nobody is watching the terminal.

Rollout dates for Codex Security Cloud. OpenAI describes Codex Security Cloud and the cleaner terminal interface as rolling out, but gives no dates for when they reach every account. The sources also give no availability dates for the Daybreak Blue model access or for building agents on AWS through Bedrock Managed Agents. It is still open whether published environments start tasks noticeably faster for teams, which is OpenAI's own claim.