How to set up agents in Claude Code
Share
Claude Code can run subagents: focused specialists you hand a job to while you keep working. Set up right, they turn one session into a small team. Set up wrong, they just burn tokens.
The basics
An agent is a scoped instruction set with its own tools and its own lane. You define it once, then call it for the jobs it's good at. The wins come from three things:
- One job per agent. A reviewer reviews. A debugger debugs. Narrow beats broad.
- The right tools per agent. A read-only reviewer shouldn't be able to write files.
- Clear hand-off. The agent returns a clean result, not a transcript.
Ten agents worth having
Planner, code reviewer, debugger, security analyzer, TDD guide, researcher, refactor cleaner, doc updater, build-error resolver, and an end-to-end tester. Each one covers a job you'd otherwise do by hand or forget to do at all.
The fast path
Writing and tuning ten agents from scratch is a project in itself. The Blubber Starter Kit ships ten customized agents already dialed in, plus the folder structure that keeps them organized. Prefer to build your own? Start from the free open-source repo.