OpenAI launches Codex app on macOS for agentic coding

A new macOS Codex app turns AI coding into a multi-agent workflow, with built-in worktrees, reusable “skills,” and scheduled automations. OpenAI is positioning it as a cleaner way to supervise longer, messier engineering tasks.

A shift from “one assistant” to “many agents” 

This week, OpenAI is placing a bet on the next productivity boost that comes from a different framework: not one coding assistant, but multiple agents working simultaneously, each operating independently enough to move quickly without interfering with the others.

Codex is a command center, not a chatbox 

On Monday, February 2, 2026, the company introduced the Codex app for macOS, describing it as a desktop interface built to manage multiple agents at once, collaborate on long-running tasks, and run work in parallel.

OpenAI’s pitch is that this isn’t just “autocomplete.” It’s closer to a command center where you can keep several tasks going across projects, review changes as they arrive, and move between threads without losing the state of each job.

The timing is also obvious. OpenAI is playing catch-up in a sector currently dominated by Anthropic’s Claude Code, where developers have already shifted toward agent-heavy workflows.

The features that make Codex feel always-on 

The headline feature is multi-agent orchestration, but the implementation is what matters. Instead of locking up your terminal, the app uses built-in worktrees, isolated copies of your repo that allow multiple agents to code in parallel without crashing into each other.

Beyond the plumbing, OpenAI is pushing “skills” to package team conventions and scripts into reusable capabilities. The announcement points to integrations for pulling design intent from Figma or updating tasks in Linear, but the real shift is scheduled runs for routine triage or failure summaries.

It suggests a future where an agent keeps working after you close the laptop, turning coding into an asynchronous loop rather than a live chat.

Security and access, without pretending risk disappears 

OpenAI says the Codex app follows a security by design approach. By default, agents are constrained to the working directory and can’t quietly escalate privileges. However, they will hit a hard stop when they need network access or sensitive permissions, forcing a manual approval, a friction point that is intentional.

Access is tied to ChatGPT plans. The app is available immediately on macOS, included for paid tiers, with limited-time access for some lower tiers and optional extra credits depending on usage.

The engine under the hood: GPT-5.2-Codex 

The app sits on top of a Codex stack that OpenAI has been steadily upgrading. In December, OpenAI introduced GPT-5.2-Codex, positioning it as a more capable agentic coding model with stronger long-horizon work through context compaction and an explicit emphasis on defensive cybersecurity.

That matters because multi-agent workflows only feel “real” when models can hold context across hundreds of steps without hallucinating imports or breaking the build.

What to watch next 

If Codex becomes the place where teams delegate maintenance chores, reviews, and migrations, the biggest competitive pressure won’t be about who writes the cleanest function. It will be about who builds the most trustworthy system for supervising agents over hours or days.

OpenAI CEO Sam Altman framed the appeal in human terms: “The models just don’t run out of dopamine,” he said during a briefing. “They keep trying, they don't run out of motivation.”

Now the question is whether developers adopt the new habit where letting several agents run, then judging the work like a lead engineer, not a typist.