Codex Reusable Cloud Environments Escalate the Coding Agent Race

What OpenAI shipped

OpenAI used its DevDay event on September 29 to change how Codex runs in the cloud. Codex, the company's software engineering agent, now works inside reusable development environments that a developer can reach from a computer, from a phone, or entirely in the cloud.110 Codex could already send tasks to remote machines. The difference is that those machines are no longer throwaway sandboxes. Environments now persist and can be configured, where previously each cloud task got a fresh, isolated setup.14

In practice, a team sets up an environment once, with its repositories, dependencies, tools and configuration, publishes it, and starts new tasks from it without rebuilding anything.59 Each task still runs in its own isolated workspace. Changes made during one task don't leak back into the shared base. Updates to a published environment affect only new tasks, and existing tasks keep their state.5 Tasks keep running while the developer's computer is asleep and can be picked up again on desktop, web or mobile.59

Environments were only one part of the release. OpenAI also shipped:

• A refreshed Codex CLI with voice control, an /agents view for handing off and tracking several tasks at once, better prompt editing, session resumption and worktree handling, and a cleaner terminal interface.147
• A code review experience in the ChatGPT desktop app. Developers can read summaries, inspect diffs and question Codex before leaving feedback on GitHub pull requests or GitLab merge requests. Automatic reviews can make a first pass while the developer is away.18
• Codex Security Cloud, which scans GitHub repositories on demand, on a schedule or when new commits land. It investigates findings, removes duplicates and prepares fixes in the cloud, and it includes access to models from OpenAI's Daybreak Blue cybersecurity initiative without a separate application.14 A community summary of the release says Security Cloud is available to eligible Pro, Business, Enterprise and Edu users on desktop and web, and that Codex CLI 0.159.0 became the stable build alongside the announcement.9

Where the coverage agrees, and what it leaves out

The reporting agrees on the facts. Every account describes the same shift from one-off remote sessions to persistent, shareable workspaces with approved settings and permissions.16710 The differences come in how much weight writers give OpenAI's claims.

The most skeptical coverage points out that faster task startup is something OpenAI says it is aiming for. The announcement includes no measurements to show it.6 Another analysis notes that OpenAI hasn't published separate pricing for reusable environments, a maximum persistence window, or independent benchmarks for Security Cloud's detection quality and false-positive rate.8 The same piece places the launch in a longer strategy. It ties the feature to OpenAI's June agreement to acquire Ona, formerly Gitpod, whose persistent environments were meant to let Codex work beyond a single device or active session.8

The more upbeat coverage presents the update as Codex turning from a personal assistant into a shared team resource.7 One commentary argues that the security scanner, not the environments, is the bigger story, because it pushes OpenAI into code governance, a space usually held by specialist security vendors.3

Our reading is that the environment feature is the foundation and the rest depends on it. Automatic reviews that run while the laptop is closed, and security scans that fire on every commit, both need a persistent, pre-approved workspace. Without one, each job would have to rebuild its own context. Security Cloud gets the attention, but the environment layer is what makes unattended operation possible. One clarification matters for buyers: the security fixes are proposed for review, and the coverage does not say they are applied automatically.6

The competitive backdrop: everyone is building the same thing

Codex's release fits a pattern across the industry. A September survey of the field called asynchronous background agents "the new normal." It listed GitHub Copilot's coding agent, Codex automations, Cursor's cloud agents, Claude Code's cloud sessions and Devin as tools that work in the background and return pull requests.11

Cursor has moved fastest in October. Its background agents have been renamed cloud agents. They still run in isolated virtual machines, clone a repository, work on a separate branch and open a pull request.14 On October 6, Cursor added remote control for agents running on a developer's own machine, operated from a phone, browser or another device.12 On October 16, it shipped interactive canvases that agents can share, plus an Environments API in beta for creating and managing agent environments in code.19 Cursor had already let teams run cloud agents on their own infrastructure, with planning and inference still handled in Cursor's cloud.19 This is the closest competitor to OpenAI's environment model. Codex packages environments as a shared, approved workspace. Cursor exposes them as programmable resources.120

The Cursor numbers also show why figures need checking. Cursor's own release note says environments support up to 300 repositories and load five times faster.19 Another report gives a ceiling of 500 repositories.20 Figures for how many of Cursor's merged pull requests come from cloud agents range from more than 35%20 to 56%.13 Both describe Cursor's own engineering team, not its customers, so treat them as directional rather than benchmarks.20

Claude Code has been shipping at a heavy pace. Between August 10 and October 20 it went from v2.1.227 to v2.1.305, 66 versions, most of them fixes.22 The October changes overlap heavily with OpenAI's agenda:

• Claude Mods arrived on October 1 and let plugins change deeper behavior.21
• Forked subagents became the default in 2.1.302.24
• Cloud sessions now keep tasks running on Anthropic's infrastructure for Pro and Max users.23
• Code Review, which reviews an organization's pull requests, became generally available on October 15 with per-repository review levels.22 A later patch fixed a bug where Code Review missed the latest commit after multiple pushes during a review.29

GitHub Copilot has taken a different route, competing on model breadth and reach. In early October it made Gemini 4 Argon generally available, added Claude Haiku 5.5, introduced free local sandboxing and Ollama model discovery, and opened a computer-use preview in Copilot CLI and its desktop app.12 Copilot CLI can now send work to local Ollama models, which cuts cloud calls on routine tasks.12 Copilot also runs async cloud agents on GitHub Actions.15 Comparisons generally find its agent mode less autonomous than Cursor's cloud agents or Codex Cloud.16

Why code review is now the battleground

All four vendors are converging on review because more agent work in the background means more output for someone to check. Codex now does a first pass on pull requests automatically.1 Anthropic's Code Review has reached general availability.22 Cursor's cloud agents can subscribe to the pull requests they open and keep working until they're done, fixing CI failures and responding to bot comments.19 Release-tracking coverage of the same October week showed Claude Code, Codex CLI, OpenCode and VS Code all shipping features to run more autonomous agents, in more places, with less supervision.24

The governance questions grow with that autonomy. Long-lived environments collect credentials, dependencies and permissions. Shared configurations only help governance if teams can audit changes, revoke access quickly and keep secrets out of files the agent can read.8 OpenAI's documentation says that sharing an environment does not grant access to another user's task, and that connected GitHub repository permissions still apply.5

The verdict

OpenAI's update is less a leap ahead than a move to close a gap. Cursor has offered cloud agents in isolated VMs since May16 and is now letting teams script and self-host those environments.19 Anthropic has persistent cloud sessions and generally available code review.2223 What sets Codex apart is the bundle: shared environments, a multi-agent CLI, review in ChatGPT, and an always-on security scanner with access to Daybreak Blue models.14

Whether that bundle wins depends on things OpenAI hasn't disclosed yet: pricing, persistence limits and independent evidence that the security tooling finds real problems without flooding teams with noise.8 For now, a cloud workspace that persists across devices is the minimum a serious coding agent needs to offer. Competition is moving to how well each vendor governs work the agent does without supervision.