docs/ide-extensions/agent/quick-start.mdx
Agent mode equips the Chat model with the tools needed to handle a wide range of coding tasks, allowing the model to make decisions and save you the work of manually finding context and performing actions.
<Accordion title="Learn how to choose the right mode" icon="lightbulb"> <Columns cols={3}> <Card title="Chat Mode" icon="comments"> *Learn and discuss without changing code.* **Mental Model:** Talking to a knowledgeable colleague
**Best For:** Explaining concepts, comparing approaches, code review discussions.
</Card>
<Card title="Plan Mode" icon="binoculars">
*Safely explore and plan with read-only tools.*
**Mental Model:** Architect surveying before renovation
**Best For:** Understanding a codebase, bug investigation, planning implementations.
</Card>
<Card title="Agent Mode" icon="hammer">
*Make actual changes with full tool access.*
**Mental Model:** Contractor executing approved blueprints
**Best For:** Implementing features, fixing bugs, running tests and commands.
</Card>
You can switch to Agent in the mode selector below the chat input box. The mode selector offers three options:
Agent mode lives within the same interface as Chat mode, so the same input is used to send messages and you can still use the same manual methods of providing context, such as @ context providers or adding highlighted code from the editor.
With Agent mode, you can provide natural language instruction and let the model do the work. As an example, you might say
Set the @typescript-eslint/naming-convention rule to "off" for all eslint configurations in this project
Agent mode will then decide which tools to use to get the job done.
By default, Agent mode will ask permission when it wants to use a tool. Click Continue to allow Agent mode to proceed with the tool call or Cancel to reject it.
You can use tool policies to exclude or make usage automatic for specific tools. See MCP Tools for more background.
Any data returned from a tool call is automatically fed back into the model as a context item. Most errors are also caught and returned, so that Agent mode can decide how to proceed.