Agent Teams: Multiple Agents, One Project

Two ways to run multiple agents

The manual way: you start several Claude Code sessions in parallel, each in its own git worktree so their changes don't collide. The coordinated way is Claude Code's Agent Teams: one session becomes the team lead, which assigns tasks and synthesizes results; teammates work independently, each with its own context window.

How a team works

The team coordinates through a shared task list: the lead creates tasks, teammates take them or claim the next available one on their own. Unlike subagents, teammates can message each other directly — for example, to challenge each other's hypotheses. Agent teams are experimental and disabled by default (as of July 2026); they must be enabled via an environment variable.

Typical roles

Proven patterns from the docs: a writer implements while a reviewer checks with fresh eyes. Or several reviewers examine the same pull request from different angles (security, performance, test coverage). Or several agents pursue competing hypotheses while debugging and try to disprove one another.

The cost warning

Each teammate is a separate Claude instance with its own context window — costs scale with team size. Anthropic's engineering blog puts multi-agent systems at roughly 15 times the tokens of a normal chat. For sequential tasks or edits to the same file, a single session is usually the better choice.