Run 10 Projects at Once: The AI Parallel Work Playbook

Garry Tan runs 10-15 parallel Claude Code sessions simultaneously. Each one implements features, reviews code, tests, or ships — independently. He checks in on the decisions that matter and lets the rest run.

This isn't a developer trick. It's a management philosophy that applies to any knowledge work.

The old bottleneck was your attention. You could run one project at a time, maybe two if you multitasked aggressively (and poorly). AI agents break that constraint: you initiate workstreams, each makes progress autonomously, and you become the decision-maker instead of the doer.

The Mental Model: CEO, Not Operator

Stop thinking of AI as a tool you use. Think of it as a team you manage.

An operator does the work: writes the code, writes the email, does the research. One thing at a time.

A CEO initiates work, checks progress, makes decisions, and redirects when needed. Ten things at a time.

You're becoming a CEO of AI agents. Your job is:

1. Define the objective clearly
2. Kick off the workstream
3. Check in at decision points
4. Approve, redirect, or kill

The Parallel Playbook: 5 Steps

Step 1: Break Work Into Independent Streams

For any large project, identify the streams that can run in parallel:

Example — launching a new product feature:

• Stream 1: Competitive research (what exists, positioning)
• Stream 2: Landing page copy and design
• Stream 3: Technical implementation
• Stream 4: Email announcement sequence
• Stream 5: Customer case study for launch day
• Stream 6: Internal training docs for support team
• Stream 7: Pricing analysis and packaging

Each stream has a clear deliverable and minimal dependencies on the others.

Step 2: Set Up Dedicated Context for Each Stream

Each stream gets its own AI context. Don't mix them in one conversation — context pollution kills quality.

For Claude: Create separate Projects, each with relevant documents and a system prompt defining the role and objective.

For Cursor: Use separate windows or Composer sessions for independent tasks.

For Claude Code: Run separate terminal sessions, each in its own branch or working directory.

Step 3: Brief Each Agent Like You'd Brief a Direct Report

Good briefing: "Research the top 5 competitors for AI contract review tools. For each, document: pricing, key features, target customer, notable customers, and their main weakness. Deliver as a comparison table with a summary of our positioning opportunities."

Bad briefing: "Do some competitive research." (Vague → bad output → you waste time redirecting.)

Spend 5 minutes on a clear brief to save 2 hours of back-and-forth.

Step 4: Run a Decision Loop, Not a Task Loop

Once streams are running, your job changes. You're not doing the work — you're making decisions:

Every 30-60 minutes, cycle through your streams:

• Stream 1 (research): "Here are the findings." → You: "Good. Now go deeper on competitor X — their pricing model seems off."
• Stream 2 (landing page): "Here are 3 headline options." → You: "Option 2, but make it more specific to enterprise buyers."
• Stream 3 (implementation): "I've drafted the API. Here's the design." → You: "Approved. Implement and write tests."

You're making 15-20 micro-decisions per hour. Each decision unblocks an agent. This is 10x faster than doing each task yourself.

Step 5: Converge and Ship

As streams complete, bring the outputs together:

• The research informs the landing page copy
• The implementation determines the technical content in the docs
• The pricing analysis shapes the email announcement

This convergence step is pure human judgment — synthesizing parallel outputs into one coherent launch. AI helps, but you're the integrator.

Real-World Parallel Work Examples

For Founders

Stream	Agent Task	Your Decision
Product	Build feature prototype	Approve UX direction
Marketing	Write 5 blog posts for launch	Pick which 3 to publish
Sales	Research and draft outreach to 50 prospects	Approve messaging
Finance	Model 3 pricing scenarios	Choose the pricing
Legal	Draft terms of service update	Review key clauses

5 streams, 5 decisions per hour, 5x the output.

For Managers

Stream	Agent Task	Your Decision
Planning	Draft Q3 roadmap	Prioritize the themes
People	Write performance review drafts	Add personal observations
Process	Audit team workflows for inefficiency	Approve changes
Reporting	Generate board update	Adjust narrative framing

For Consultants

Stream	Agent Task	Your Decision
Analysis	Market sizing for 3 segments	Validate assumptions
Deliverable	Draft 40-page strategy deck	Refine key recommendations
Proposal	Write 3 new client proposals	Personalize and send
Content	Write thought leadership articles	Add your unique perspective

The Attention Management System

Parallel work fails without a system for tracking where things stand.

Simple approach: A running document (or Claude Project) that tracks:

• Stream name
• Current status (waiting on AI / waiting on my decision / complete)
• Next decision needed
• Priority (ship-blocking vs. nice-to-have)

Review this document every hour. Handle the decisions that unblock the most downstream work first.

The Compound Effect

Working in parallel doesn't just save time — it produces better results. When you do competitive research before writing the landing page, the landing page is better. When you build the prototype before writing the sales email, the email is more concrete.

Parallel streams create information flow between workstreams that sequential work can't match. You have more context for every decision because multiple investigations are happening simultaneously.

Start Small

Don't try to run 10 streams on day one. Start with 2-3 parallel AI conversations on your next project. Get comfortable with the decision loop — brief, check in, decide, redirect. Once that feels natural, add more streams.

The goal isn't maximum parallelism. It's optimal use of your judgment — spending your time on the decisions that matter most while AI handles the execution.

Join AI Builder Club to learn multi-agent workflows from builders running them in production.