Claude Code Subagents and Hooks

You spawn multiple subagents. They start cold. They burn 85% of your tokens. Again.

You thought subagents would make Claude Code feel like a dev team. Instead, one vague prompt fans out, your plan limit melts, the trace looks fine, and the PR is still wrong. Meanwhile, the operators ahead of you are running multi-agent stacks overnight: five hooks watching the run, one kill switch aborting the runaway, and a settings.json that drops the token bill 30% before the next invoice hurts.

This book ships the working stack: a multi-agent Claude Code team, five hooks, worktree isolation, and one issue-to-PR workflow that turns Claude Code into your own dev team. By Chapter 4, you have a .claude/agents/ folder running on your repo. By Chapter 6, hooks are logging every worker. By Chapter 8, one skill, multiple subagents, and one hook are wired into a real issue-to-PR pipeline. By Chapter 10, a kill-switch hook catches the runaway before it eats the session. By Chapter 12, your token bill is down 30% and your PR output is up 2x.

This is not another "subagents cookbook." Recipe dumps tell you to install 100 specialists. This book teaches the discipline that actually ships: a small multi-agent team, not a sprawling specialist library. A planner, a fixer, and a reviewer. Self-contained briefings. Hook logs you can read. Worktrees that keep parallel workers from clobbering your repo. A Goldilocks Test that retires every subagent that is not earning its slot.

Here is what you will build:

1. A working multi-agent stack wired to a planner: investigator, fixer, reviewer.
2. Five hooks that turn silent subagent runs into a readable trace.
3. A kill-switch hook that aborts runaway subagents before they cost you $50.
4. A composition workflow: one skill + multiple subagents + one hook in a real issue-to-PR pipeline.
5. A worktree-based parallel-fanout script that does not clobber your repo.
6. A ~/.claude/settings.json cost-control template that cuts your token bill 30%.
7. The four-field prompt template that stops cold subagents from guessing.
8. The Goldilocks Test that retires bad subagents before they bloat your stack.
9. A 30-day calendar from chat-box habit to multi-agent Claude Code workflow.



Five new "subagents cookbook" books drop on Amazon every week with stale commands, recycled patterns, and 500-page recipe dumps that ship nothing. This one is different: 190 pages of operator discipline, a multi-agent stack you run by Chapter 4, and the kill-switch hook that can pay for the book in your first week of token savings. Stop burning tokens on blind agents. Scroll up and grab your copy.