Chatting with the agent

Agent mode is Startbase's autonomous execution mode. When you give it a task, it takes ownership end to end: understanding your intent, exploring the codebase for context, applying changes across files, and resolving issues that appear during development. You can follow along step by step and review the results before moving on.

• Implement features or changes once an approach is decided
• Fix bugs or errors end to end
• Refactor code across multiple files
• Apply coordinated changes across frontend, backend, and configuration
• Debug issues that arise during implementation
• Inspect logs and preview output to identify problems
• Verify results before finishing

Agent mode is active by default. Describe what you want and send your message. Clear requirements and constraints produce better results.

Be specific about what you want
"Add a contact form with name, email, and message fields. Validate email format and show error messages inline."

Describe expected behavior
"When users click 'Add to Cart', show a success message and update the cart count in the header. Store cart items in local storage."

Include constraints when relevant
"Add a dark mode toggle. Use the existing Tailwind dark: variants. Don't change the current color palette, just add dark equivalents."

While the agent works, the chat shows a live working row with what it is doing right now. Finished work collapses into grouped summaries ("Made 6 actions") you can expand to see every file edit, command, and check. Longer tasks are organized into phases you can collapse and expand.

• Stop ends the current run; everything already applied stays in place.
• If the agent proposes a plan first, it appears as a plan card. Open View to read the full plan before it executes.
• Closing the tab does not lose anything: reopen the project and the chat reattaches to the running turn.

You can keep typing while the agent works. New messages queue up and run in order when the current turn finishes. Queued messages stay editable: change or remove them any time before they start.

• Rate any reply with the feedback control under it; your signal improves the product.
• Report a reply that went wrong from the message menu, with an optional note.
• The enhance button (wand) rewrites your draft prompt into a sharper one before sending. Your intent is preserved; the structure improves.

The model picker in the composer selects which AI model powers the agent and how much it thinks. See Models and thinking levels for a full guide.