Agents

What are agents?

When you describe a plugin to Claude, it doesn't do everything alone. It sends specialized helpers — called agents — to handle different parts of the job. Each agent is good at one thing, and they work together to deliver your finished plugin.

Think of it like a small team: one person reads the blueprint, another writes the code, and a third checks the work.

How agents build your plugin

Here's what happens behind the scenes when you say something like "Build a plugin that renames selected layers to lowercase":

1. Understanding your idea

Claude starts by figuring out exactly what you need. It asks clarifying questions if anything is ambiguous — just like a designer would before starting a project.

2. Exploring the codebase

An explorer agent scans the project files to understand the current state — what code already exists, what components are available, and where changes need to happen.

3. Writing the code

An implementer agent writes the actual plugin code. It creates the Figma API logic, the React UI, and the message types that connect them. It follows the patterns already established in the template.

4. Reviewing the work

A reviewer agent checks the code for mistakes, missing pieces, and quality issues. If something isn't right, the implementer fixes it before you ever see it.

5. Building and testing

Claude runs the build process to make sure everything compiles without errors. If the build fails, it diagnoses and fixes the problem automatically.

You don't need to manage agents

Agents work automatically — you never need to tell Claude which agent to use or how to coordinate them. Just describe what you want and let Claude handle the rest.

The only time you'll notice agents is in Claude's status messages, where you might see things like "Exploring codebase..." or "Reviewing implementation..." as different helpers take turns.

What makes agents useful

• Fresh perspective — Each agent starts with a clean slate, so mistakes don't carry over
• Specialized knowledge — The reviewer catches issues the implementer might miss
• Parallel work — Some agents can work at the same time, making things faster
• Quality checks — Your code gets reviewed before it's presented to you