Agents

Nimbus runs coding agents inside the sandbox so they can inspect files, execute commands, preview changes, and report concrete verification results.

Agent turns

An agent turn starts when you send a prompt. The worker records the turn, sends it to the sandbox bridge, streams progress back to the app, and stores the final assistant response.

Good prompts

Give agents enough context to act without guessing:

• Describe the desired behavior.
• Mention relevant files, screens, or commands.
• Say whether the change should be committed, pushed, or only left in the worktree.
• Ask for specific verification when correctness matters.

Tool access

Agents can use shell commands, file edits, browser automation, and desktop inspection depending on the task. For UI work, browser validation is preferred because it checks the rendered behavior rather than just the source.

Completion

A turn is complete when the assistant response is recorded and the UI no longer shows the agent as working. If a turn appears stuck, inspect the session logs and bridge health before retrying.

Repository safety

Agents should preserve unrelated worktree changes. When committing, they should include only the files required for the task and leave unrelated untracked files alone.