Claude Code Agent Teams: What It Does and Where It Stops

Agent Teams is Claude Code's built-in mechanism for running multiple coordinated sessions instead of one. Per Anthropic's documentation, it is experimental and gated behind an environment flag: you opt in, and Claude Code gains the ability to spawn and coordinate teammate sessions from a lead session. It was introduced in early 2026 and got a simplified setup in mid-2026, which is the version current as of this writing.

The important framing: this is Anthropic productizing a workflow people were already assembling by hand, several terminal tabs, each running Claude Code in its own git worktree, with a human shuttling context between them. Agent Teams moves the shuttling into the tool: the sessions know about each other, can message each other, and pull from a shared task list rather than waiting for you to copy-paste instructions.

An agent team has one lead session, the one you talk to, and a set of teammates it spawns and coordinates, with a documented range of up to 16 sessions. The lead decomposes your request into tasks, hands them out, and integrates results; teammates work their assignments and report back. Four properties make this more than a fan-out of subagents:

If you have used Claude Code's subagents, the distinction matters: subagents are ephemeral helpers inside one session, results flow back to the parent and vanish. Teammates are peer sessions that persist for the life of the team and coordinate laterally. (The FAQ below covers this in more depth.)

Agent Teams shines where work is genuinely parallel and loosely coupled. Two shapes come up repeatedly in Anthropic's own framing and in community coverage:

Fan-out research. “Investigate these five subsystems and report back” is the ideal team task: each teammate reads deeply in its own context window, nothing needs to merge, and the lead synthesizes findings that would have overflowed a single session's context. This is the multi-session version of the context-isolation argument, each investigator gets a clean window.

Parallel independent tasks in one repo. Three bug fixes in three different modules, or a feature plus its docs plus a test-coverage pass, tasks that do not share files, can run simultaneously on separate worktrees, with the shared task list keeping assignments straight. The broader design question of decomposing work so agents do not collide is the subject of multi-agent coding workflow; Agent Teams is one engine you can run such a workflow on.

Where it strains: tightly coupled work. If every task touches the same files, teammates serialize on each other or produce conflicting changes, and a single strong session (or one agent with subagents) is usually faster.

None of these are criticisms, Agent Teams is an experimental feature doing exactly what it says, but they define where it stops today:

If you mostly need “several sessions without losing my mind”, the boundaries may never bite; the spectrum of lighter and heavier approaches is mapped in managing multiple Claude Code sessions.

The natural question: if Claude Code can coordinate its own sessions, do you still need an orchestration tool? They solve different layers of the problem, and the fair comparison looks like this:

The honest conclusion is that they compose. Agent Teams is an execution engine: excellent at short-lived, coordinated bursts of parallel Claude sessions. An orchestrator like AIDEN is a workflow layer: persistent stories, an enforced spec gate, worktrees, and PRs per card. Some users run Agent Teams inside an AIDEN story, the story provides the spec, branch, and review gate, and the team fans out within it. Where each layer of that stack belongs is the subject of Claude Code orchestration.