"Agents" is the biggest buzzword in AI right now, and the honest answer to "how do I create one in Claude" is simpler than most people expect: you don't build a separate agent app. You get Claude to work agentically -- to take a goal, plan the steps, use your tools, and act on its own until the job is done.

What "an agent" actually means in Claude

There's no "New Agent" button to hunt for. Agentic work in Claude shows up in two forms:

• Claude acting on a goal. Give it a multi-step objective and the tools to do it, and it plans and executes -- reads files, pulls from your apps, drafts, saves -- without you steering every step.
• Subagents. For bigger jobs, Claude spins up focused workers, each with its own clean context, and can run several in parallel.

The simplest way to "make an agent": give Claude a goal and tools

An agent is really just Claude with a clear objective, access to your tools, and permission to take the steps. You describe the outcome, not the keystrokes:

Connect the tools it needs, and Claude runs the whole sequence. Put it on a schedule and it does that on its own every week -- which is about as close to "an agent" as most people actually want.

Subagents: run focused work in parallel

When a job is big or repetitive, Claude can hand slices to subagents -- separate workers with their own clean context that go do the work and report back:

One request, several workers, clean results back to combine.

Reusable agent "types"

You can also define reusable agent types -- a named worker with its own instructions and tools, like a "researcher" or a "reviewer" -- and delegate to it by name whenever that work comes up. It's the agent version of a saved skill.

Where agents really pay off

The magic isn't one clever agent. It's chaining agentic steps into systems -- research feeds a draft, a draft feeds a review, a review feeds a send -- so a whole workflow runs end to end, often on a schedule, while you do something else. That's how one person starts operating like a team.