AI/ML Development

The Parallel Claude Workflow

Running Multiple AI Agents Like It's 1995 Multi-Tasking (But Actually Good)

From the engineer who remembers when running two programs simultaneously meant your Commodore 64 would freeze. Now I orchestrate multiple AI agents in parallel terminals, each building different features across isolated git worktrees.

The Problem Space

You know what takes me back? The first time I tried to run multiple programs simultaneously on my Commodore 64. Spoiler alert: it didn't work. The machine would just... freeze. Or crash. Or produce some beautiful, chaotic screen artifacts that looked more like modern generative art than anything functional.

Fast forward forty-odd years, and here I am orchestrating multiple AI agents in parallel terminals, each one working on different features across isolated git worktrees. The technology has changed, but that same fundamental challenge remains: how do you make multiple "things" work at the same time without them stepping on each other's toes?

I'll be honest—when I first heard about developers running multiple Claude Code instances simultaneously, I was skeptical. Having worked with AI coding assistants for a while now, I know the bottleneck isn't usually the AI's speed—it's my review bandwidth. These things generate code faster than I can think about whether it's any good.

The Serial Workflow Problem

I'd kick off Claude to refactor some authentication logic, and while it was churning away, I'd just... sit there. Waiting. Context-switching to Slack. Grabbing coffee. Coming back to find it done, reviewing it, then starting the next task.

The serial approach felt like running a single-core processor when I had multiple cores available.

Three Parallel Patterns

There are essentially three approaches to running multiple Claude instances in parallel. Each has its trade-offs, and the "right" choice depends on your project structure and tolerance for operational overhead.

Pattern Comparison

Choose the right approach for your workflow

💡 Recommendation

For most professional workflows, Git Worktrees offer the best balance of isolation and practicality. Add tmux when you need session persistence or remote development capabilities. Start with 2 parallel agents maximum until you understand your review bandwidth.

Pattern Two: Git Worktrees

Now we're talking. Git worktrees are one of those features that's been around since 2015 but somehow still feels like secret knowledge. (Kind of like how SIMD instructions existed for years before most developers actually used them.)

The concept is elegant: instead of cloning your repo multiple times, git lets you check out different branches into different directories, all sharing the same underlying git object store. It's like having multiple working directories that are all "views" into the same repository history.

Setup Commands

# Your main repo
cd ~/project
# Create worktrees for parallel work
git worktree add ../project-auth feature-auth
git worktree add ../project-payment feature-payment
git worktree add ../project-refactor refactor-api

Why This Works Better

Each worktree has its own filesystem. Agent 1 can modify ../project-auth/api/auth.go while Agent 2 modifies ../project-payment/api/auth.go — and they're different files in different directories, even though they represent the same logical file on different branches.

Git also prevents you from checking out the same branch twice. Try to create two worktrees on feature-auth and git will stop you. This is the kind of safety mechanism I appreciate—it's like having a mutex at the git level.

Git Worktree Isolation

Click a worktree to simulate Claude working in parallel

~/projectmain

Main repository - shared git object store

Quick Reference

git worktree add ../project-auth feature-auth
cd ../project-auth && claude

The Operational Overhead

Here's what that ZX Spectrum kid in me wishes someone had told me upfront: worktrees require bootstrapping. Each one needs:

  • Dependencies installed (npm install, poetry install, whatever your stack uses)
  • Configuration files that aren't in git (.env files, local secrets)
  • Build artifacts if your project has a compilation step
  • Database migrations run if you're working on schema changes

For a Node.js project with a hefty node_modules directory, this can mean significant disk space. Yes, git efficiently shares objects, but those dependencies? They're duplicated.

Pattern Three: Tmux + Automation

If you're old enough to remember GNU Screen (or still use it), tmux will feel familiar. If you're younger and never had to deal with unreliable SSH connections that would kill your entire dev session... well, lucky you.

For parallel Claude workflows, tmux offers something valuable: persistent, detachable sessions. You can launch three Claude instances in three tmux sessions, detach from all of them, close your terminal, come back hours later, and they're all still running.

The Power of Persistence

Tmux sessions survive SSH disconnects, terminal closures, and laptop suspends. For remote development or long-running AI tasks, this resilience is invaluable. Your agents keep working even when you're not watching.

Tmux Session Manager

Persistent, detachable Claude instances

Active Sessions

claude-auth
detached

Auth Refactor

claude-payment
detached

Payment API

claude-api
detached

API Restructure

No session attached
$ tmux attach -t <session-name>

Tmux Commands

tmux new -s claude-auth
tmux attach -t claude-auth
tmux ls # list sessions

Automation Scripts

The real power comes when you automate the entire setup. Here's a battle-tested script pattern:

parallel-claude-setup.sh
#!/bin/bash
PROJECT_ROOT=~/my-project
FEATURES=("auth-refactor" "payment-api" "api-restructure")

for feature in "${FEATURES[@]}"; do
    # Create worktree
    git worktree add "${PROJECT_ROOT}-${feature}" "feature-${feature}"

    # Bootstrap
    cd "${PROJECT_ROOT}-${feature}"
    npm install
    cp ../.env .env

    # Launch in detached tmux session
    tmux new-session -d -s "claude-${feature}" \
        "cd ${PROJECT_ROOT}-${feature} && claude"
done

echo "Launched ${#FEATURES[@]} Claude instances"
tmux list-sessions

Run once, get three parallel agents working on three different features. Each in its own tmux session, each with its own worktree, each completely isolated.

The Token Reality

Here's the part nobody mentions in the glossy "parallelize everything!" tutorials: token consumption is real, and it scales non-linearly.

Running three Claude instances isn't 3× the tokens—it can be more, because each agent explores its own context, makes its own mistakes, backtracks its own dead ends. I've hit my Claude Pro subscription limits multiple times doing this.

Token Consumption Monitor

Real-time tracking across parallel agents

Total Usage
57,617 / 100,00057.6%
Agent 1: Auth

15,234

tokens consumed

45 tok/s
Agent 2: Payment

23,451

tokens consumed

67 tok/s
Agent 3: API

18,932

tokens consumed

52 tok/s

💡 Key Insight

Running 3 parallel agents isn't 3× the tokens—it can be more. Each agent explores its own context, makes its own mistakes, and backtracks independently. Monitor your usage closely and consider your review bandwidth as the true bottleneck.

The Realities Nobody Mentions

Review Bandwidth Is Still Your Bottleneck

Remember my earlier skepticism? It's still valid. You can parallelize implementation, but you can't parallelize your own critical thinking. Three agents producing code simultaneously means three sets of changes to review, understand, and integrate.

Context Switching Has Cognitive Cost

Jumping between three different feature contexts while each Claude asks you questions is like trying to debug three different production incidents simultaneously. Possible? Yes. Pleasant? Not particularly.

Merge Conflicts Are Still Merge Conflicts

Worktrees prevent agents from literally overwriting each other's files during implementation. They don't prevent merge conflicts when you try to integrate all those feature branches. If Agent 1 and Agent 2 both refactored the same module, you're still going to spend time resolving that manually.

Sweet Spot: Two Agents

I've found my sweet spot is two agents for active work, maybe a third for long-running research or refactoring tasks I can review asynchronously. Beyond that, the coordination overhead exceeds the parallelization benefits.

What I'd Recommend to Past Me

1.
Start with two parallel agents, not five.

Learn the workflow with manageable complexity.

2.
Script the worktree bootstrap process early.

That 5 minutes of bash scripting saves you 30 minutes every time you spawn a new agent.

3.
Use tmux from the start.

Even if you think you don't need persistent sessions, you do. Your SSH connection will drop. Your laptop will suspend. Tmux saves you from starting over.

4.
Monitor your token usage.

Set up alerts or just check the dashboard regularly. Surprise subscription limits are no fun.

5.
Accept that review is your bottleneck.

Don't try to scale beyond your review capacity. Two well-reviewed implementations are better than five half-reviewed ones.

The Meta-Question: Should You Do This?

Here's the thing—just because you can run multiple AI agents in parallel doesn't mean you should. It's a power tool, and like any power tool, it can cut through work efficiently or it can cause spectacular disasters.

✓ Valuable For

  • • Large features that can be cleanly decomposed
  • • Refactoring work across different modules
  • • Exploratory research on multiple approaches
  • • Long-running tasks while you focus on high-priority work

✗ Not Valuable For

  • • Closely related features with merge conflicts
  • • Work where you need tight feedback loops
  • • Complex features with evolving architecture
  • • Teams where coordination overhead exceeds benefits

Your mileage will vary based on project structure, team size, and personal work style. The current state of parallel Claude workflows is like shell scripting in 1995—functional but primitive.

We're figuring this out together,
one parallel workflow at a time.

Explore More AI & Development Workflows

Dive deeper into MLOps, automation, and developer productivity patterns.

Browse All Articles
The Parallel Claude Workflow: Running Multiple AI Agents | ASleekGeek