Before reaching for a team, Anthropic's own guidance is to check whether a lighter option does the job: subagents inside one session, or separate sessions that you run yourself and that pass findings between them (Source 1). The difference is in who talks to whom.

The table is Anthropic's comparison (Source 1), condensed. Notice that it gives no numbers for cost, and neither does this chapter: the documentation says only that teams use significantly more tokens and that usage scales with the number of active teammates.

Agent teams cannot be run the way the other samples in this guide were, and section 03 explains why. What can be run non-interactively is the step below the team: two subagents from Chapter 20 working at the same time on different files. I asked a small, cheap main model (haiku) to coordinate, and let the two agents' own files choose their models:

What the log shows, and what it does not:

• Parallel: the main session issued both Agent calls in a row before either reviewer had finished, and the table interleaves the reviewers' own calls. Their results arrived separately.
• Mixed models: the log lists both claude-haiku-5-5 and claude-sonnet-5-5. I did not trace which step used which.
• A messy edge: the main session first answered "still running in the background", then produced a combined answer after the reviewers reported. The log contains three result events, and the cost shown is the sum of all three. The second repeats the first, so the figure is slightly high. If you script around claude -p and subagents, expect more than one result and do not assume the first is final.
• Checked: both findings were verified against the code and by measurement (see the author's check). The combined answer was properly cautious about what the reviewers had not run.

The facts, from Anthropic's page:

• It is off by default. Set the environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to 1 in your shell or in settings.json under env. Without it no team is set up and Claude does not propose teammates.
• Enabling it changes ordinary delegation. While it is on, a subagent that Claude gives a name to launches as a teammate, so a team can form even when you did not ask for one. Setting the variable to 0 turns that back off.
• You describe the team in plain words. The lead session creates the shared task list, spawns teammates and synthesises results. You can ask for specific names and models ("use Sonnet for each teammate").
• Two display modes. In-process (the default; all teammates inside your terminal, with arrow keys and Enter to open a teammate's transcript) or split panes, which need tmux or iTerm2 and are not supported in VS Code's integrated terminal, Windows Terminal or Ghostty. The teammateMode setting and an experimental --teammate-mode flag choose between them.
• Roles can be reused. You can name a subagent definition when spawning a teammate, so a sql-reviewer file works both as a delegated subagent and as a teammate. The definition's tools and model apply and its body is added to the teammate's instructions. Its skills are not applied, and mcpServers applies only to split-pane teammates.

Anthropic's page recommends starting with research and review, which have clear boundaries and avoid the file-conflict problems of parallel coding, and says the strongest cases are research and review, new modules, competing-hypothesis debugging and work that spans layers. These three prompts follow that advice for bookshelf. They are the author's, written for this chapter and not executed. They are patterns, not recorded results.

The documentation's reasoning for this shape is that a single investigator tends to stop at the first plausible explanation, while several investigators trying to disprove each other are less anchored (Source 1). That is the documented mechanism, not something shown here.

The hook rows are the practical lever. They are how you stop a teammate marking a task done while the build is red. The hooks lab in Chapter 22 covers the whole contract; the short version: exit code 2 is the "block and tell Claude" signal in Claude Code's hook system.

The documented limitations are worth reading as a checklist before you start, because several are about losing work:

• No session resumption for in-process teammates. /resume and /rewind do not restore them; the lead may try to message teammates that no longer exist.
• Task status can lag. Teammates sometimes do not mark tasks complete, which blocks dependent tasks. You may need to nudge or update status by hand.
• One team per session, no nested teams, fixed lead. Teammates cannot spawn teams, and a teammate cannot become the lead.
• The lead can finish early or start doing the work itself. The page suggests telling it to wait for teammates.
• Split panes need tmux or iTerm2, and orphaned tmux sessions can remain after a session ends.

On size: the documentation suggests starting with three to five teammates and about five or six tasks per teammate, and says token cost scales linearly with the number of active teammates. A team is a bet that parallel exploration is worth that multiple. For routine work, a single session or a subagent is the better bet, which is the same answer the comparison table gave.