← deeps.dev
Agent Teams
← deeps.dev
01 — Architecture

What makes up an agent team?

Think of it this way: instead of one Claude doing everything, you now have a team lead that delegates work to independent teammates. Each teammate is a full Claude Code session — its own context window, its own tools, its own workspace.

They stay coordinated through two shared resources: a task list (what needs doing) and a mailbox (how they talk to each other).

Team Lead
Teammates
Mailbox
Task List
Key insight: Teammates don't inherit the lead's conversation history. They load project context (CLAUDE.md, MCP servers) fresh — so you need to give them enough context in the spawn prompt.

Where things live on disk

Teams and tasks are stored locally. Team config lives at ~/.claude/teams/{name}/config.json with a members array. Tasks live at ~/.claude/tasks/{name}/. Teammates can read the config to discover other team members.

02 — Comparison

Subagents or agent teams?

Both let you parallelize work, but they solve different problems. The question to ask: do your workers need to talk to each other?

Subagents are like sending assistants on errands — they go, do the work, and bring back a summary. Agent teams are like a war room — everyone's working simultaneously and can challenge, inform, or redirect each other in real time.

SubagentsAgent Teams
TopologyStar — report back onlyMesh — peer-to-peer
ContextResults return to callerFully independent windows
CoordinationMain agent manages allShared task list + self-claim
Token costLowerHigher (N separate contexts)
Best forFocused, quick tasksComplex collaborative work
Rule of thumb: If you just need results back, use subagents. If teammates need to share findings, debate, or coordinate on their own — use agent teams.
03 — Tasks

How tasks flow through the team

The lead breaks work into tasks and adds them to a shared list. Each task has a state: pending, in progress, or completed. Tasks can depend on other tasks — a blocked task won't become claimable until its dependency finishes.

When a teammate finishes their current work, they look at the list and self-claim the next available task. Claiming uses a file lock to prevent two teammates from grabbing the same work.

Pending
In Progress
Completed

Watch the tasks move...

Tasks start in Pending. Teammates claim them (file-locked). Blocked tasks wait. Once a dependency completes, downstream tasks automatically unblock.

Sweet spot: Aim for 5–6 tasks per teammate. Too few and there's nothing to pick up when they finish early. Too many and the list becomes noise.
04 — Communication

How teammates talk to each other

There are three ways agents communicate. The important thing: all delivery is automatic. No one needs to poll for messages. When a teammate sends something, it arrives immediately.

Direct Message

Send a message to one specific teammate. Use this for targeted instructions, follow-up questions, or redirecting someone's approach. It's the most common communication pattern.

Watch the cost: Broadcast sends to every teammate — token usage scales with team size. Use it sparingly, only for global instructions or status updates.
05 — Display

Two ways to see your team

You can run all teammates inside your existing terminal, or give each one its own pane. Both work — it's a matter of preference and what your terminal supports.

claude — agent-teams

In-Process Mode (default)

All teammates run inside your terminal. Use Shift+Up/Down to select a teammate, Enter to view their session, Escape to interrupt. Ctrl+T toggles the task list. Works everywhere — no extra setup.

06 — Decision Guide

Should you use agent teams?

Agent teams add coordination overhead and use significantly more tokens. They're worth it when parallel exploration adds real value — but for sequential work or same-file edits, a single session is faster and cheaper.

Walk through this decision tree. Hover each node for a best-practice tip.

Quick rules

07 — Use Cases

Where teams really shine

These are the scenarios where agent teams deliver the most value. Each card shows the communication pattern — watch how data flows differently depending on the task type.

Getting started: If you're new to agent teams, start with research or review tasks — they have clear boundaries and don't require writing code. This shows the value of parallel exploration without the coordination challenges of parallel implementation.
Sources & Attribution

Where this comes from

This visual guide is an educational companion to Anthropic's official Claude Code documentation. All technical content, architecture details, and feature descriptions are sourced directly from the docs below. The goal is to make the concepts easier to grasp through interactive visuals — not to replace the original material.

If you're looking for the full, authoritative reference, always start with the official docs.

Primary source

Orchestrate teams of Claude Code sessions
Official Claude Code documentation — covers architecture, configuration, communication, display modes, use cases, best practices, and known limitations.
code.claude.com/docs/en/agent-teams

Related documentation

Note: This page is an independent educational resource, not affiliated with or endorsed by Anthropic. It's built to help people understand agent teams visually. For the most up-to-date and complete information, refer to the official documentation.
Summary for this section
Agent Teams let you coordinate multiple Claude Code instances. A Team Lead spawns Teammates (each with independent context), coordinates via a shared Task List and Mailbox. Best for parallel research, code review, debugging, and cross-layer work.