8 setups · worktrees, subagents and overnight bots
Run multiple Claude Code agents in parallel: 8 real setups
Last updated · by Thomas, editor · How we verify
These are setups people run when one coding agent at a time is too slow: several Claude Code or Codex sessions side by side, a lead agent that splits the work into pieces, or a bot that hands tasks out overnight.
What they share: a person stays in the loop, approving plans, permission prompts or merges. Every recipe links to the post, repo or doc it came from.
Which setup fits how you work
How each setup splits the work, what keeps agents from colliding, and where you step in.
The Claude Code team's habits: parallel worktrees, plan first, skills
Several Claude sessions at once, one task each. Kept apart by: a git worktree per session.
⏱ Setup: 15 min✋ A second session reviews each plan; you pick which fixes to applyRun several coding agents on a kanban board
One Claude Code or Codex agent per task. Kept apart by: the app makes a worktree per agent.
⏱ Setup: 20 min✋ You answer every permission prompt from the Needs Input columnGet notified when your coding agent needs you
You run the sessions; a hook flags whichever one is waiting. Kept apart by: the folder name in each notification.
⏱ Setup: 10 min✋ You approve the hook settings, then each promptSplit a big task into a DAG of small subagents
A Claude Code plugin that breaks work into atomic tasks with dependencies and gives each subagent only the context it needs, in parallel worktrees.
⏱ Setup: 20 min✋ You approve the graph before dispatch; merge after checks passDone-when lists and a two-try stop rule
One lead agent (Claude Code or Codex) splits each task into bounded pieces for other models in separate worktrees, checks every report against a done-when list, and scales review to how hard a change is to undo.
⏱ Setup: 1 hour to set up✋ Critical changes wait for an independent review; you mergeInstall specialist subagents from a community library
Named specialist subagents you call inside one session. Kept apart by: each subagent file sets its own tools.
⏱ Setup: 10 min✋ You read each file and its permissions before installingA Grok Bot front door to your live Claude Code sessions
A Grok Bot types each task into the right live session. Kept apart by: one herdr workspace per project, across 3 machines.
⏱ Setup: Not stated (his full guide: plan on a few sessions)✋ No merge or force push without approval; you review each PRNight shift: labeled GitHub bugs → reviewed draft PRs by morning
A Chief bot plus Triage, Builder, Reviewer and Release bots. Kept apart by: cloud agents in separate environments; one owner per issue.
⏱ Setup: Half a day to set up✋ Draft PRs only; you read the morning report and merge
Details come from each recipe's source. Setup time is shown only where the recipe states it.
Sessions side by side
You run each session yourself, one task and one worktree each, and only open the one that needs you.
The Claude Code team's habits: parallel worktrees, plan first, skills
The creator of Claude Code shared his team's habits: run 3–5 sessions in parallel worktrees, start complex work in plan mode, and turn anything you do daily into a skill.
Run several coding agents on a kanban board
Each agent gets its own git worktree and terminal; a board sorts them into Idle, Working and Needs Input so you only look at the ones blocked on you.
Get notified when your coding agent needs you
He had Claude Code configure itself so he gets a notification whenever it is waiting on input or a permission.
A lead agent splits the work
One agent plans, hands bounded pieces to others and checks what comes back.
Split a big task into a DAG of small subagents
A Claude Code plugin that breaks work into atomic tasks with dependencies and gives each subagent only the context it needs, in parallel worktrees.
Done-when lists and a two-try stop rule
One lead agent (Claude Code or Codex) splits each task into bounded pieces for other models in separate worktrees, checks every report against a done-when list, and scales review to how hard a change is to undo.
Install specialist subagents from a community library
A large curated collection (161+ when we checked) of Claude Code subagent definitions across 10 categories that you can drop into a project.
Bots hand out the tasks
A bot routes work to coding agents, often overnight, and reports back.
A Grok Bot front door to your live Claude Code sessions
One Grok Bot's only job is passing tasks to his Claude Code agent. It finds the right live session in herdr on one of his 3 machines, types the task in, and reports back with the machine, repo, branch and PR link.
Night shift: labeled GitHub bugs → reviewed draft PRs by morning
A Chief bot and four specialists turn labeled GitHub bugs into reproduced, fixed, independently reviewed draft PRs overnight, with a morning report that separates done from attempted.
How to choose
- Start with two or three sessions. The Claude Code team runs 3 to 5, and the recipe warns that too many sessions overload your review capacity. (recipe)
- Give every agent its own worktree, and plan around shared files. Worktrees keep checkouts apart, but parallel tasks editing the same file still conflict, so split the work by files. (recipe)
- Make the waiting agent find you. A notification hook or a Needs Input column means you only open the session that's blocked, instead of scrolling through all of them. (recipe)
- Write the finish line first. A done-when list a program can check, plus a stop after two failed tries, keeps a worker from looping and keeps its report honest. (recipe)
- Decide what may merge without you. The done-when setup lets only routine, easy-to-revert changes merge after CI. The night shift opens draft PRs and blocks merges with branch protection, not just the prompt. (recipe)
Common questions
Can you run multiple Claude Code agents at once?
Yes. The setups here run each session in its own git worktree so agents don't edit the same files. The Claude Code team's habit is to run several sessions in parallel worktrees, each on a separate task, and start complex work in plan mode.
How many parallel agents should I run?
The Claude Code team runs 3 to 5 sessions. The limit people hit is review, not the agents: the same recipe lists too many sessions overloading your review capacity, and shared dev databases still cause interference between worktrees.
Subagents or separate sessions?
Subagents work inside one session: Orchestra gives each one only its objective, spec sections and files to touch, and the community library installs specialists you call by name. Separate sessions in worktrees suit independent tasks you review one at a time. Some setups here use both.
Does this work with Codex too?
Yes. Agent Navigator supports Claude Code and Codex CLI, and the done-when setup leads from either one and hands pieces to other models in separate worktrees.
Should parallel agents merge their own work?
Mostly not here. The night shift opens draft PRs and blocks merges with branch protection, and the front door can't merge or force push without approval. The done-when setup lets routine, easy-to-revert changes merge once CI passes, while critical ones wait for an independent review and a human merge.
Get the next money recipe. Free.
New recipes every week, plus the top 10 agent setups pack. One email, no spam.
Unsubscribe anytime. Privacy
More guides: Make money with AI agents · AI voice agent business · AI agent agency ideas · Claude Code use cases · Grok Bot use cases · Claude Code vs Codex · AI agents for small business · AI email agents · AI code review agents · Autonomous agents that made money