Codex for (Almost) Everything - OpenAI Turns Its Coding Agent Into a Full Workspace

OpenAI just dropped the biggest Codex update since launch, and the headline isn’t really about code anymore. Codex can now operate your computer, browse the web, generate images, remember your preferences, and schedule its own future work. The name still says “coding agent” but the ambition clearly says “development operating system.”

What landed

The update is dense. Here’s what matters:

Background computer use. Codex can now see your screen, click, and type with its own cursor - running in parallel with your actual work. Multiple agents can operate simultaneously without getting in each other’s way. This is the same computer-use paradigm Anthropic shipped with Claude, but OpenAI is framing it around iteration: testing frontends, clicking through apps that don’t have APIs, working across your actual desktop environment rather than just the terminal.

In-app browser. You can now comment directly on web pages and have Codex act on those instructions. Right now it’s positioned for localhost development - iterating on UIs and games. The plan is to expand it to full browser control, which would make it a direct competitor to browser-use agents from smaller players.

Image generation. Codex now uses gpt-image-1.5 to generate and iterate on visuals. Combined with screenshots and code, it’s targeting the design-to-code workflow: mockups, product concepts, game assets. This is a clear shot at tools like v0 and Lovable that have been eating the prototyping market.

90+ plugins. Atlassian Rovo, CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Neon, Render, and more. The plugin count is becoming a moat. If your entire stack is connected, switching away gets harder every week.

Memory. Codex now remembers context from previous sessions - preferences, corrections, hard-won knowledge about your codebase. This is the kind of thing that sounds incremental but is actually transformative. The difference between an agent that starts from zero every time and one that builds up institutional knowledge is the difference between a junior hire and a senior one.

Automations and scheduling. Codex can now schedule future work, wake itself up to continue long-running tasks, and reuse existing conversation threads. This is the “set it and forget it” pattern - landing open PRs, following up on Slack threads, staying on top of fast-moving conversations across tools.

Proactive suggestions. Using context from your projects, connected plugins, and memory, Codex now proposes what you should work on next. Open comments in Google Docs, relevant Slack threads, stale tasks - it surfaces a prioritised list. Whether this is helpful or annoying depends entirely on execution, but the direction is clear.

The strategic read

Step back and the pattern is obvious. OpenAI is building the development equivalent of Microsoft Office - a single workspace that owns the entire lifecycle, from ideation to code to review to deployment. Every update pushes Codex further from “smart terminal” and closer to “where you do all your work.”

The computer-use feature is the most telling addition. It’s an admission that the API-only approach has limits. Some apps don’t have APIs. Some workflows require clicking through a UI. Some testing needs actual visual verification. Rather than waiting for the world to build better APIs, Codex is just operating the interface directly. It’s messy, but it works.

Memory and automations together solve the real problem with AI coding assistants: context loss. Every developer who has used Copilot or Cursor or Claude Code has experienced the frustration of re-explaining your project to a fresh session. If Codex can genuinely accumulate and apply institutional knowledge across sessions, the productivity compounding is real.

What’s still rough

Computer use is macOS only at launch, with EU and UK availability “coming soon.” If you’re on Linux or Windows, you’re waiting. The Mac-first approach mirrors Anthropic’s Claude launch strategy, but it still leaves a chunk of developers out.

Memory is in preview. OpenAI’s track record with memory features in ChatGPT suggests this will be a slow rollout with some rough edges. The question isn’t whether it works on day one, but whether it degrades gracefully when it gets things wrong. A coding agent that confidently remembers the wrong thing is worse than one that forgets.

The scope is getting unwieldy. Codex is trying to be a coding assistant, a browser, a design tool, a project manager, and a scheduler all at once. There’s a real risk of becoming a jack-of-all-trades that masters none. The plugins help, but integration breadth without depth is just a dashboard.

Why this matters

The coding agent space is consolidating fast. Anthropic has Claude Code with deep terminal integration. Cursor owns the editor experience. GitHub Copilot has the install base. OpenAI’s play is breadth - be the one tool that touches every part of your workflow, from JIRA ticket to deployed code.

Whether that works depends on execution. The best development environments are opinionated and fast, not all-encompassing and sluggish. But if OpenAI can make the integrations feel native rather than bolted-on, the 3 million weekly users give them a distribution advantage that’s hard to match.

The competitive moat isn’t the model anymore. It’s the workflow. And OpenAI just made a very clear statement about where they think that workflow lives.