The built-in agents

Now that you know what to delegate, here’s the shortcut: you usually don’t have to build the helper. Claude Code comes with a few specialists already on staff.

Each built-in agent has its own tools and its own instructions, tuned for one kind of job. You just point it at the right work.

It’s like a company that already has a researcher, an architect, and a generalist on payroll. You don’t write a fresh job description each time — you say “this one’s for the researcher” and hand it over.

The regulars

• Explore — the finder. Read-only and fast, built to locate things: where is the login logic, which files import this, where does this value come from. It won’t change anything; it goes, looks, and reports back.
• Plan — the architect. For designing an approach before any code is written: how should we add support for X? It returns a step-by-step plan and weighs trade-offs, instead of diving in.
• general-purpose — the generalist. For multi-step jobs that don’t fit a specialist — research that mixes searching, reading, and reasoning into one answer.

You can also define your own agents — a job description you write once and reuse — but the custom side belongs to Level 11. For now, the regulars cover most of it.

Why this beats “just prompt harder”

Each agent comes pre-loaded with the right tools and the right instructions. Explore, for instance, has no editing tools at all — which is exactly why it’s safe and fast to send off searching. It physically can’t wander off and change something.

Picking the agent that fits the job is usually a bigger lever than wording your prompt more cleverly.

Right specialist beats better speech. When a job has an obvious shape — finding, planning, general research — reach for the agent built for that shape first.

What’s next

Subagents are help you summon for one job. The next multiplier is different: knowledge Claude reaches for on its own, the moment a task calls for it.