Guide: Agents

Claude Cowork · Feature primer

Overview

An agent is Claude working toward a goal with some independence. Instead of you directing every step in a back-and-forth, you hand the agent a defined task and it works through the steps — reading, deciding, using tools — and returns a result. Sometimes the main Claude you’re talking to spins up a sub-agent: a focused worker it delegates a slice of the job to, so the slice gets done thoroughly without cluttering the main thread.

The mental model is delegation. You’re less a driver and more a manager handing off a piece of work.

Agents vs. Projects

These get confused because both make Claude more capable, but they answer different questions.

• A Project is where context lives. It stores your files, instructions, and memory so Claude knows your business. See the Projects guide.
• An agent is who does the work. It takes a task and executes it across multiple steps.

They compose. You can run an agent inside a Project so the worker has both a job to do and the context to do it well. One is the knowledge; the other is the labor.

When Work Gets Delegated

You don’t always set agents up by hand. On a complex task, Claude may delegate parts of it to sub-agents on its own — sending a focused research question to one worker, a verification pass to another — and then bring the results together. This keeps each piece clean and lets independent work happen in parallel.

You’ll see this most on bigger jobs: “research these ten companies,” “review this whole codebase,” “produce this multi-part report.” The work fans out, each part is handled, and the answers come back combined.

When You Actually Need One

Agents shine when a task is multi-step, somewhat independent, and benefits from Claude pushing through to a result rather than checking with you at every turn:

• Research that means visiting many sources and synthesizing them.
• A repeatable job you’d rather kick off and review at the end than babysit.
• Work that fans out into independent pieces — several accounts, several documents, several checks.

For a quick question or a single edit, you don’t need an agent — a normal conversation is faster. Reach for the agent model when the work is big enough that delegating beats steering.

Building Your Own

Most people never need to build an agent from scratch — the ones bundled in plugins, or the delegation Claude does automatically, cover the common cases. For teams that do want custom multi-agent workflows, there’s a developer path (the Agent SDK) for wiring up purpose-built agents. That’s a deeper topic than this primer; what matters here is recognizing when a job wants a worker rather than a conversation.

When to Reach for an Agent

• The task has many steps and you’d rather review the outcome than guide each move.
• The work splits naturally into independent pieces that can run at once.
• You want Claude to carry a job to completion, using tools along the way, with you checking the result.

If you’d be tempted to walk away and come back to a finished result, that’s the work an agent is for.