Skip to content

Code

Collaboration is a pairing session — you drive, the agent assists. Unlike planning (where you consult on architecture) or work dispatch (where agents lead end-to-end), a collaborate session is open-ended. You direct the work; the agent follows. There is no structured procedure.

  • A repository with Cate set up (quickstart)
  • A description of what you’re working on

Open the Collaborate command from the dashboard. Provide a brief description of what you’re working on — this becomes the title of the tracking issue and the basis for the branch name.

The agent creates a lightweight tracking issue on your board (placed in In Progress) and proposes a branch name. It asks you to confirm or adjust. Once you confirm, it checks out the branch and the session begins.

The session is free-form. There’s no phased structure — you direct, the agent follows. Some things you might do:

  • Ask the agent to implement a function or component
  • Point it at a failing test and ask it to debug
  • Request a refactor of a specific module
  • Ask it to write tests for code you’ve been working on
  • Explore an unfamiliar part of the codebase together

The agent stays responsive, not proactive. It doesn’t restructure your work, set agendas, or propose next steps unless you ask.

The agent suggests committing at natural save points: “We’ve got the tests passing — want me to commit before we move on?” You can also ask it to commit at any time. Commits use your project’s format and reference the tracking issue.

When working on UI changes, the agent may offer to produce ASCII mockups to align on layout before coding.

When you’re done — say “done,” “ship it,” “wrap up,” “let’s create a PR,” or similar — the agent walks through a closing sequence. Each step is an offer, not a mandate.

If your project has guardrails defined, the agent offers to run the quality gates before creating a PR. If any required gate fails, it tells you and asks how to proceed.

If there are uncommitted changes, the agent offers to commit them.

If you say yes, the agent pushes the branch and opens a PR with a summary of the session’s changes.

If you say no, your changes stay on the branch for you to push later.

If a PR was created, the tracking issue moves directly to Human Review — skipping the automated agent review step. Since you were present for the entire session, there’s no need for an agent to review the code before you see it.

The issue is in Human Review. You (or a teammate) review the PR and merge when ready. See Review features for the human review workflow.