OpenAI just released the Codex app for macOS. The interesting part isn’t “better autocomplete” — it’s workflow: you can run multiple agents in parallel, keep them separated by project/thread, and (critically) use Git worktrees so agents can work on the same repo without trampling each other.
Source: OpenAI announcement — Introducing the Codex app
https://openai.com/index/introducing-the-codex-app/
What the Codex app gives you (practically)
From OpenAI’s announcement, the Codex app includes:- Threads organised by project — each agent runs in its own thread, so context and diffs stay contained.
- Built-in git worktrees — multiple agents can work on the same repo without branch/checkout chaos.
- Diff review in the thread — review changes where the work happened; comment, iterate, then ship.
- Skills — packaged workflows (instructions + scripts + resources) so agent work is repeatable.
- Automations — scheduled runs for the boring-but-important stuff (CI babysitting, issue triage, release briefs), with results landing in a review queue.
Why worktrees are the real feature
If you’ve tried running multiple agents against one repo, you’ve probably seen some mix of:- lockfile churn
- accidental branch switching
- partial changes overwriting each other
- “it worked on the agent branch” but your local checkout is now inconsistent
How I’d actually use this
A realistic setup:- Agent A: fix CI failures / build issues
- Agent B: implement a feature
- Agent C: write tests + do a review pass
TL;DR
- Codex app for macOS = multi-agent UI + project threads + built-in worktrees
- Worktrees are the practical win: parallel agents without branch hell
- Skills/Automations are promising, but the day-to-day value is isolation + review