agent-teams

think in teams, not prompts.

---

your single agent has a ceiling. you've felt it. the context window fills with noise from earlier tasks. independent work runs one thing at a time. you hit the limit, type /clear, and start over.

agent teams break through by giving each worker its own clean context window. one session leads. the others work in parallel, each focused on a single piece. unlike subagents — which can only report back to the parent — teammates talk to *each other*. they share findings, challenge assumptions, coordinate directly.

it's the difference between sending people on separate errands and putting a team in a room.

---

// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

one env var. restart claude code.

optional but recommended: install tmux so each teammate gets its own visible pane. without it, everything runs in one terminal — cycle through with Shift+Down.

requires claude code v2.1.32+. check with claude --version.

---

each teammate is a full claude code session with its own context window. tokens scale linearly.

1 agent = 1×. 3 agents ≈ 3–5×.

the best recipe: plan first in plan mode (~10k tokens, cheap). hand the plan to a team for execution (~100k tokens, focused). going in blind can burn 500k+ tokens in the wrong direction.

reserve teams for work that justifies the spend.

---

[try-it/](try-it/) — three prompts to copy-paste tonight. hands on keyboards.

[go-deeper/](go-deeper/) — the patterns that separate burning tokens from getting results.

[field-notes/](field-notes/) — real projects, real costs, real failures. the stories you'll remember.

---

is the task complex enough to need more than one session? if no — single session. fast and cheap.

can the subtasks run independently in parallel? if no — subagents. sequential focus.

do the workers need to coordinate directly