What happens at four key steps

Ask modeA read-only mode for asking questions about a codebase without changing files; the safe way to explore unfamiliar or legacy code. Press Enter for the full definition. makes zero file changes. It lets the user inspect a legacy repository before changing it.

Use a prompt such as "Explain this repository. Use diagrams where they help." The result should help a senior engineer check the agent's understanding of the code.

Plan modeA mode that makes no edits: it researches the codebase and produces an editable plan you review before any code changes. Press Enter for the full definition. asks design and scope questions, then pauses when the request conflicts with the repo.

Stage a real gap in which the PRDProduct Requirements Document. The spec describing what to build, for whom and why, before implementation starts. Increasingly the artifact you hand an AI agent as the source of truth for a feature. Press Enter for the full definition. assumes authenticated users, but the repository has no authentication. The agent asks whether to build authentication or ship a local-first V1. That proves it challenges the plan instead of blindly executing it.

A cloud agent can open a browser, test the change, capture a walk-through video plus screenshots and correct its own console errors.

It attaches those artifacts, type checks and test results to the pull request for human review. That evidence makes the agent's work reviewable. Cursor has reported that cloud agents create about 30 to 40% of its internal pull requests. Verify that figure before quoting it.

Forked-repo CI checks appear inside the cloud agent UI, where you can ask why a check failed and watch the agent rerun, diagnose and fix it in front of the room.

A first-party automation can start this when CI breaks. The broken check becomes proof that the gate fires while a human keeps the merge key.