Claude Agent Teams: How One Prompt Now Deploys an Entire AI Workforce

What Agent Teams Actually Are

Standard AI tools work sequentially: you ask, it answers, you ask again. Claude Agent Teams breaks that model entirely. You define a complex goal, and Claude spawns multiple fully independent instances — called teammates — to pursue it simultaneously. One instance acts as the team lead: it decomposes the project, assigns each workstream to a specialist teammate with its own context window and tools, and synthesizes the final output. Teammates communicate and share findings directly with each other without you coordinating the handoffs. Unlike subagents, which only report back to the main agent, you can also message individual teammates directly to redirect their approach mid-execution. This is not AI working faster. It is AI working at a different organizational scale.

The Lead and Teammate Architecture: How Coordination Works

The coordination model is deliberate. When a teammate finishes planning its approach, it sends a plan approval request to the lead. The lead reviews it and either approves or rejects with specific feedback. If rejected, the teammate revises and resubmits before touching anything. Once approved, execution begins autonomously. The lead makes these approval decisions without asking you — but you can influence its judgment by embedding criteria in your initial prompt: 'only approve plans that include test coverage,' or 'reject any plan that modifies the database schema.' This means you are not managing a team. You are writing the governance rules for a team that manages itself. The quality gate is built into the architecture, not into your attention.

Real-World Proof: 100,000 Lines, 16 Agents, Zero Supervision

Anthropic's own research team ran one of the most demanding tests of this architecture: they tasked 16 agent teammates with writing a Rust-based C compiler from scratch capable of compiling the Linux kernel. The result, across nearly 2,000 Claude Code sessions: a 100,000-line compiler that runs on x86, ARM, and RISC-V. The researcher's key insight was that the system works best when you design the environment for the agents rather than directing them step by step. Clear test suites, progress files that agents update continuously, and task-locking mechanisms to prevent two agents from solving the same problem simultaneously. The agents oriented themselves using these structures without human guidance. Netflix has independently deployed multiagent orchestration for its platform team in production.

Dreaming: The Layer That Makes Agents Self-Improve

On May 7, 2026, Anthropic added Dreaming to Claude Managed Agents — a scheduled process that makes agents self-improving over time. After sessions complete, Dreaming reviews the execution history, extracts behavioral patterns, identifies what worked and what did not, and updates the agent's memory store automatically. The next time the agent runs, it starts from a more informed baseline. You control how much autonomy Dreaming has: it can update memory automatically, or you can review proposed changes before they are applied. Together with persistent memory across sessions, this creates a compounding system: agents do not just execute tasks, they become progressively better at the specific workflows you run. The gap between a newly configured agent and one that has been running for a month becomes measurable.

Prompt

# CLAUDE AGENT TEAMS — HOW TO ACTIVATE AND USE

# Requirements: Claude Code v2.1.32 or later
# Check your version:
claude --version

# ─── STEP 1: ENABLE AGENT TEAMS ───
# In settings.json, add:
{ "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
# Or set the environment variable in your shell:
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

# ─── STEP 2: DEFINE A GOAL THAT BENEFITS FROM PARALLEL WORK ───
# Agent teams work best when work can be divided with minimal dependencies.
# Bad fit:  Sequential tasks where step 2 needs step 1's output
# Good fit: Parallel investigation, separate modules, competing hypotheses

# ─── STEP 3: DESCRIBE THE TEAM STRUCTURE IN YOUR PROMPT ───
"Create an agent team to analyze this product from three angles:
- Teammate 1: Market positioning and competitive landscape
- Teammate 2: Technical feasibility and architecture risks
- Teammate 3: Devil's advocate — find every reason this will fail
Have them share findings with each other and produce a synthesis."

# ─── STEP 4: INTERACT WITH INDIVIDUAL TEAMMATES ───
# In-process: Shift+Down to cycle through teammates
# Type to send a message directly to a specific teammate
# Press Enter to view their session, Escape to interrupt

# ─── STEP 5: THE LEAD HANDLES PLAN APPROVAL ───
# Each teammate submits a plan to the lead before executing
# The lead approves, rejects with feedback, or requests revision
# To influence the lead's judgment, add criteria to your prompt:
# "Only approve plans that include test coverage"
# "Reject any plan that modifies the database schema"

# ─── STRONG USE CASES ───
# Research: multiple teammates investigate different aspects simultaneously
# New features: each teammate owns a separate module
# Debugging: teammates test competing hypotheses in parallel
# Cross-layer work: frontend / backend / tests each owned by one teammate