When you send a prompt, you’re not talking to a chatbot that spits out code snippets — you’re directing an agent that works inside a real, running copy of your project.
Your workspace
Every session runs in its own isolated cloud workspace: a full copy of your project with its dependencies installed and the app actually running. That’s why the preview beside the chat is live — it’s your real app, not a mockup.
Maddie · just now
Add a low-stock warning to the dashboard — flag anything at or under its reorder point
Vision
9 steps
Added a Low stock section — items at or under their reorder point get an amber badge.
Guardrails · all 6 checks passed
Items
1,284
Low stock
7
Orders today
32
Workspaces are elastic. Leave for a while and the session parks itself (your work is checkpointed to GitHub first); come back and it resumes right where you were. Starting fresh is fast too — Vision pre-warms a workspace while you’re still typing your first prompt.
The turn loop
Each prompt kicks off a turn:
- Rules first — the agent reads your guardrails before touching anything.
- Plan and build — it explores the codebase, makes changes, and runs commands. Tool cards in the chat show each step, expandable if you want the detail.
- Ask when unsure — if it needs a decision or a credential, it asks in chat instead of guessing. Sensitive commands come back as approval cards; needed integrations appear as one-click connect cards.
- Report — the preview updates as changes land, and the turn ends with a summary of what changed.
Everything is a version
Every prompt creates a named, restorable version with a checkpoint behind it. Browse them in the history drawer, preview any of them safely, and restore with one click — the full story is in Rollbacks.
Chats and context
A session can hold multiple chats, each with its own agent context. Start a new chat for a new topic — it keeps context sharp, and old chats remain browsable. For genuinely parallel work, run multiple sessions or hand a batch of tasks to the Manager — see Building in parallel.
What the agent sees from you
Prompts can carry more than text: paste images (screenshots, mockups), attach markdown files (specs, notes), or point at a specific element of the running app via Visual edit. The more precise the input, the better the output — more on that in Improve accuracy.