Prompt Generator

Configure your multi-agent workflow and generate a ready-to-paste prompt for Claude Code's orchestrator agent.

1. What are your agents doing?
Required
2. Team Configuration
Configure
3. Agent Behavior
Optional

Template Library

Pre-built prompt templates for common multi-agent workflows. Click any template to expand and copy.

CLAUDE.md Builder

Generate a complete CLAUDE.md system prompt for multi-agent orchestration. Customize the rules your orchestrator follows.

Project Context
Orchestration Rules

Core Concepts

The foundational ideas behind multi-agent orchestration in Claude Code.

"The true constraint of agentic engineering is twofold: the tools we have available and our ability to prompt engineer and context engineer the outcomes we're looking for."

Andy Devdan

The Three Pillars

01

Multi-Agent Orchestration

A primary orchestrator agent creates task lists and spawns specialized sub-agents. Each sub-agent gets its own context window, model, and session ID. They work in parallel on focused tasks, communicate results back via send_message, then get deleted to force clean context resets. The orchestrator compiles results while keeping its own context lean.

02

Multi-Agent Observability

You can't improve what you can't see. A custom observability system captures every session start/end, tool call, task update, and inter-agent message. Running Claude Code inside T-Mux provides visual panes for each sub-agent. Andy's system shows swim lanes, event streams, and per-agent drill-downs — 60+ tool calls per minute across teams of 8 agents.

03

Agent Sandboxes

Isolated execution environments (like E2B) where agents do work without jeopardizing your local machine. Andy ran 24 sandboxes simultaneously. Pattern: rapid prototype in sandbox → if you like it, copy locally. Sandboxes live ~12 hours. Also works with dedicated hardware (Mac Mini) for privacy-sensitive workflows. This is the "scale compute" part of the equation.

The Core Four Framework

Everything in agentic engineering boils down to four elements. Every new feature is just a new combination:

📦
Context
🧠
Model
✍️
Prompt
🔧
Tools

Agent Team Lifecycle

🏗
Create Team
team_create
→
📝
Create Tasks
task_create ×N
→
🚀
Spawn Agents
Parallel launch
→
⚡
Work
Independent ctx
→
💬
Communicate
send_message
→
🧹
Delete Team
Context reset

Key Insights

Context Efficiency: 8 sub-agents explored 8 full-stack codebases and the orchestrator only used 31% of its context window.
Delete = Reset: Deleting agents after completion forces good context engineering — fresh context prevents rot.
Ad Hoc Teams: Don't just run one team — fire off new agent teams for follow-up work, fixes, and refinements.
Trigger Phrases: "Build a new agent team" are information-dense keywords that activate specific Claude Code tools.
Engineer ≠ Replaced: Engineers are the best positioned to use agentic technology. Knowing what's under the hood is the advantage.

Tool Reference

All new Claude Code multi-agent orchestration tools with descriptions and usage patterns.

Team Management

team_create
Initialize a new agent team. The orchestrator calls this first to set up the collaboration space.
New
team_delete
Shutdown and cleanup an entire team. Forces context reset. Always call when work is done.
New
Task
Kick off a sub-agent in a parallel process. Pre-existing tool, now integrated with team workflows.
Existing

Task Management

task_create
Define a specific work unit and assign it to a sub-agent. Each task should be atomic and focused.
New
task_list
View all active tasks across the team. The orchestrator uses this to track progress.
New
task_get
Check the status and details of an individual task. Useful for monitoring specific agents.
New
task_update
Modify task parameters, status, or reassign. Used for iterative workflows.
New

Communications

send_message
Inter-agent messaging — the most critical new tool. Sub-agents use this to report results back to the orchestrator. Enables coordination without context sharing.
Critical
⚠️ Important Pattern

Each sub-agent has its own independent context window. They do NOT inherit the orchestrator's context. This means every sub-agent needs to independently run any skills or /commands it needs. Always tell sub-agents which skills to load in your task description.

Decision Guide

Not every task needs a full agent team. Use this guide to decide the right approach.

Setup Guide

Step-by-step instructions to enable multi-agent orchestration in Claude Code.

⚠️ Multi-agent orchestration requires the experimental agent teams flag. This feature is currently in preview and may change.

Install T-Mux (required for pane visualization)

T-Mux gives you visual panes for each sub-agent. Without it, you can't see agents working in parallel.

brew install tmux

On Linux: sudo apt install tmux — On Windows: use WSL

Start a T-Mux session

All Claude Code agent teams run inside T-Mux. Start a new session first.

tmux

Enable the experimental agent teams flag

This environment variable unlocks the multi-agent orchestration tools.

export CLAUDE_CODE_AGENT_TEAMS=1

Launch Claude Code

Start Claude Code inside the T-Mux session. It will now have access to team management tools.

claude

Verify with a test prompt

Send a simple orchestration prompt to confirm everything is working.

Build a new agent team. Create 2 agents. Have each list the files in the current directory and report back.

T-Mux Navigation Cheat Sheet

Key bindings you'll need while agents are running:

Ctrl+B, ←/→ — Switch panes
Ctrl+B, [ — Scroll mode
Ctrl+B, z — Fullscreen pane
Esc — Exit scroll mode
All-in-One Setup Script
Run this in your terminal
#!/bin/bash # Multi-Agent Orchestration Quick Setup # Start tmux session named "agents" tmux new-session -d -s agents # Enable agent teams feature tmux send-keys -t agents 'export CLAUDE_CODE_AGENT_TEAMS=1' Enter # Launch Claude Code tmux send-keys -t agents 'claude' Enter # Attach to session tmux attach -t agents # You're now ready to use multi-agent orchestration! # Try: "Build a new agent team for each directory in this project"

Cost Estimator

Estimate API costs for multi-agent workflows. Prices based on current Anthropic API rates.

Configure Your Workflow
$0.00
Estimated Total
0
Est. Total Tokens
0 min
Est. Wall Time
Pricing Reference (per 1M tokens)
Opus 4.6
Input: $15 / Output: $75
Sonnet 4.5
Input: $3 / Output: $15
Haiku 4.5
Input: $0.80 / Output: $4

* Estimates are approximate. Actual costs depend on prompt length, output complexity, caching, and tool call overhead. Extended thinking may increase costs 2-3x for Opus/Sonnet.