If you searched for "run parallel Codexes," you are probably past the first-agent stage. One Codex session in one terminal is fine. The trouble starts when you have several terminal tabs and windows open, and you cannot quickly tell which directory, branch, or task each session owns.
Then the same-repo problem shows up. Two agents can edit the same checkout, overwrite local state, or leave you with changes that are hard to separate into clean reviews.
The practical answer is worktrees plus a good UI. Worktrees give each task its own checkout and branch. Conductor gives those worktrees a Mac app around them: chats, terminals, run scripts, diffs, checks, pull requests, and archive flow.
Use one workspace per independent Codex task
Conductor's unit of independent work is a workspace. Each workspace is backed by a git worktree, gets its own branch and working directory, and keeps its terminal, diff, checks, and pull request path separate.
Use separate workspaces when each Codex task can become its own branch, review, or pull request. This is the common path for independent features, bug fixes, issue fanout, and experiments.
Use multiple chats in one workspace when the Codex sessions need the same branch and current code state, such as one session implementing while another reviews the same diff.
Split sessions by task
A single terminal session works for one task. When work can ship separately, move it into its own workspace.
Use worktrees for isolation
Each workspace gets its own branch and working tree, so Codex is not editing the same checkout as another agent.
Keep runs attached to the branch
Use the workspace terminal or Run button so tests and servers run against the branch Codex changed.
Review from one UI
Compare each Codex diff, track checks, merge the branches worth keeping, and archive the rest.
For the full step-by-step workflow, see Run multiple Codex sessions in parallel.
Watch the workflow
Use these walkthroughs to see the terminal-to-workspace problem and the Conductor workflow in context.
They cover:
- why several Codex terminal sessions become hard to track across tabs, windows, directories, and branches
- how worktree-backed workspaces isolate independent tasks in the same repository
- how Conductor keeps each task's chat, terminal, local run state, diff, checks, pull request, and archive flow visible
- how Codex and Claude Code can run in parallel when you want to compare plans or split work across agents
Managing multiple agent sessions with Conductor
Choose when agents should use separate workspaces or share one workspace.
Conductor CEO Charlie Holtz walks through his AI coding setup
See how Conductor fits into a real parallel agent workflow.
Where native Codex worktrees fit
OpenAI documents Codex across the app, CLI, IDE, and cloud surfaces. The Codex app also documents worktrees for isolating agent work, and the Codex CLI runs locally from your terminal.
Conductor is useful when Codex should run inside the same repo workflow you use for Claude Code: reusable workspace setup, copied local files, run scripts, terminals, diff review, checks, and pull request handoff.
Read the official Codex app worktrees docs and Codex CLI docs for the native Codex model. Use Conductor when you want parallel Codex sessions managed as reviewable workspaces.
Set up the repo before you fan out
Parallel Codex sessions work best when a new workspace can become runnable without manual repair. Put static gitignored files in Files to copy. Put dependency installation, generated files, symlinks, local database setup, and other repeatable preparation in setup scripts.
If several workspaces need local dev servers at once, use CONDUCTOR_PORT so each workspace gets a separate port range.
Before you start five agents
Prepare the repo once, then fan out safely.
- Give every agent a task that can finish independently.
- Copy required `.env` files or configure Files to copy before fanout.
- Move generated setup into setup scripts instead of doing it by hand.
- Use `CONDUCTOR_PORT` when several workspaces need local servers.
- Avoid assigning two agents the same broad file-heavy refactor.
- Review, merge, or archive one workspace at a time.
Common patterns
For issue fanout, create one workspace per issue, give each Codex session a narrow prompt, and review the resulting branches separately.
For implementation plus review, keep both sessions in one workspace if the reviewer needs to comment on the same in-progress diff.
For mixed-agent work, run Codex and Claude Code in separate workspaces when their tasks can ship separately. Use one workspace when they need the same branch and latest local state.