Guide · Agentic AI · as of 7 November 2026

How to run several AI coding agents at once, without them breaking each other’s work.

By , Managing Director, BRev Digital ·

On 4 November Meta bought Conductor, a small company whose Mac app does one thing: it lets a developer run several AI coding agents side by side, each in its own copy of the code, and then review and merge what they made. A big company paying for that tells you where software work is going. One agent at a time is already the old way. We have worked like this for months. When we counted on 7 November, our main project had 65 separate working copies on one machine in Dubai, and 319 pull requests had been merged into it in the previous 30 days, 67 of them on 4 November, the day of the deal. Here is what Conductor did, what happens to its users now, how to set up parallel agents with Claude Code, and the collisions that copies of the code do not prevent.

What Meta bought

A desk for many agents.

Conductor’s own description is short: create parallel Claude Code, Codex and Cursor agents in isolated workspaces, see at a glance what each is working on, then review and merge their changes. It did not make a model. It made the place where a person supervises several models at once. Meta confirmed the purchase to TechCrunch and gave no terms. Conductor’s changelog of 4 November sets the dates its users need. Conductor Cloud goes read-only on 13 November and shuts down on 20 November. The local Mac app keeps getting new models and critical bug fixes until early 2027. A cloud workspace can be moved to your Mac with Transfer to local, and Settings, Account, Export data downloads everything. If your team used the cloud version, that is a job for next week, not next month.

Doing it without Conductor

One agent, one working copy.

The idea underneath Conductor is a git feature called a worktree: a second folder with its own files and its own branch, sharing the same history as the first. Two agents editing the same folder overwrite each other within minutes. Two agents in two worktrees cannot touch each other’s files at all. Claude Code builds this in. Start a session with claude --worktree and a name, and it creates the folder under .claude/worktrees at the root of the project, on a new branch called worktree- plus the name. Run the same command with another name in a second terminal and you have two agents in parallel. A subagent can get its own worktree too: add isolation: worktree to its definition. When you leave a session, Claude Code removes a clean worktree and asks before deleting one that has work in it. Anthropic’s documentation is honest about the limit. A worktree is a fresh checkout, so each one needs its own install, and files that git ignores, like a .env file, are missing unless you list them in a .worktreeinclude file. Everything else that is shared stays shared.

How our agents learn from their mistakes

What still collides

each one happened to us

Separate folders. Shared everything else.

  • The git stash is one stack for every worktree. An agent that pops the top entry can pick up another agent’s half-finished work. Ours label every entry and pop only the one they can name.
  • Staging by folder sweeps in strays. An agent that staged a whole directory carried two untracked files into merged pull requests, one of them a 541 KB PDF. Ours stage file by file, and a commit hook refuses the sweep.
  • A stale copy of main. Each worktree only knows the main branch as of its last fetch. Branching from a stale copy built work on a tree missing two merged changes, four times in one week. Ours fetch before every new branch.
  • Guessing a number. Pull request numbers are handed out in order, so a guessed one belongs to a neighbour, and its green checks read as your pass. Ours ask git which pull request belongs to the branch.
  • Database migrations. Two agents that each pick the next migration number by looking at the folder pick the same one. Ours take numbers from one script, under a lock.
  • Ports. Two agents that both start the app on the default port fight over it. Ours pass a port per lane, by hand.

How we merge 300 changes a month

The hard part is the queue.

Writing code in parallel turned out to be the easy half. The hard half is everything after: the checks, the merge, the database change and the deploy, done in order, by many agents at once. Until 27 October a person did that ordering, which was me, and pull requests, migrations and deploys were missed while they waited for me. Now one script does it. Every agent ships through the same command, and the command takes a lock before anything that must happen one at a time: a lock for claiming a migration number (55 numbers claimed so far), a lock for applying a change to the production database, and a lock held from bringing a branch up to date until it is merged. Each lock is a folder in the git directory every worktree shares, created in a single step that either fully happens or does not happen at all, so two agents can never both hold it. A second agent waits in the queue instead of racing. One more number, because parallel agents cost money. When we measured a week of our usage in October, subagents were 66% of it, and 89 of them had started on the most expensive model because nobody had set one. Parallel is cheap to start and easy to overspend on. Set the model for each agent on purpose.

Custom software development in Dubai

How to start with two agents this week

Two agents first. Then the queue.

  • Give each agent its own worktree. With Claude Code, claude --worktree <name> in each terminal. With other agents, git worktree add ../project-<name> -b <name>, then start the agent in that folder.
  • Add .claude/worktrees/ to .gitignore, and list the ignored files each copy needs, like .env, in .worktreeinclude. Then run the install once in each worktree.
  • Write the shared-state rules where every agent reads them: stage files by name, fetch before branching, never pop the stash blind, get the pull request number from the branch.
  • Put one lock in front of each thing that must happen one at a time: migration numbers, database changes, merges to main. A script beats a rule, because a script cannot forget.
  • Choose the model per agent. A search or a lookup does not need the largest model. Measure a week of usage before you add a third agent.
  • Keep a person on the merge until the queue is proven. The agents can write in parallel from day one. Let them merge on their own only after the checks catch what a person would.

Why this matters to us

Our software is built this way, every day.

We are an AI lab in Dubai. The apps we build for UAE businesses, from audit software to a construction suite, are written by AI coding agents working in parallel and checked by people and by other agents. The lesson of the last six months is simple to say and was not simple to learn: more agents only helps if the work they finish can reach production in order.

AI agents, built in Dubai

Questions

Questions about parallel AI agents.

  • How do I run multiple Claude Code agents at once?

    Start each one in its own git worktree. Run claude --worktree with a different name in each terminal, and Claude Code creates a separate folder under .claude/worktrees on its own branch. Each agent then edits its own copy of the code, and you merge the branches when the work is reviewed.

  • Can I run multiple Claude Code sessions in the same folder?

    You can, but they will edit the same files and overwrite each other. Reading in parallel is fine; writing in parallel needs a separate worktree for each session. Also keep shared state in mind: the git stash, the staging area and dev server ports are shared even between worktrees.

  • What is Conductor, and what happens now that Meta bought it?

    Conductor is a Mac app for running Claude Code, Codex and Cursor agents in parallel, each in an isolated workspace, with a view to review and merge their changes. Meta announced the purchase on 4 November 2026. Conductor Cloud goes read-only on 13 November and shuts down on 20 November 2026; the local app gets new models and critical fixes until early 2027.

  • What is a git worktree?

    A second working folder for the same repository, with its own files and its own branch, sharing the same history. It lets two people or two agents work on two branches at once without switching back and forth, and git worktree list shows every one.

  • How many AI coding agents can I run at once?

    The limit is rarely the machine. It is how fast the finished work can be checked and merged, and how much usage you want to pay for. Start with two, put a queue in front of merges and database changes, measure a week of usage, then add more.

  • Do parallel agents cause merge conflicts?

    Less than you would expect, because most tasks touch different files. The worse problems are the quiet ones: two agents picking the same migration number, a branch built on a stale copy of main, or one agent reading another’s test results. A lock or a single shipping script prevents those.

Sources

read 7 November 2026

Check it at the source.

  1. Conductor changelog: Conductor is joining Meta (4 November 2026, the dates for users)
  2. TechCrunch: Meta acquires AI coding startup Conductor (4 November 2026)
  3. conductor.build (what the app does)
  4. Claude Code docs: Run parallel sessions with worktrees (--worktree, .worktreeinclude, subagent isolation)
  5. Git documentation: git worktree (the underlying feature)

A question?

Running more than one agent?

Tell us how your team uses AI coding agents today. We will tell you honestly what would collide if you ran them in parallel, and what it takes to make the merge queue hold.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com