This page walks one complete loop: state the task, watch it execute, review the result.
1. Choose a workspace
Pick the workspace that contains your project. Bi treats it as a boundary: every file operation stays inside it, and paths outside are refused.
2. State a task you can accept
Instead of “take a look at this project”, give it a goal you can judge. For example:
Read the entry files under src, map the module dependencies, and write the result to DEPENDENCIES.md.
A concrete goal makes it easier for the agent to know when it is done — and easier for you to accept the result.
3. Watch it work
The conversation shows what the agent is doing in real time: which files it reads, which commands it runs, which content it changes. The first write or command asks for your approval.
If read-only steps interrupt you too often, you can auto-approve low-risk actions in Settings. See Operations that need approval.
4. Review the artifacts
Generated documents, images and code appear in the artifacts panel, ready to preview. Terminal output and the change list for this session live there too — see Artifacts panel.
5. Roll back if you do not like it
Key points keep checkpoints. If a step goes off track, return to that checkpoint and let the agent try again instead of restarting the conversation. See Checkpoints and rollback.
Next