Chat is great for a quick change. But some work is a goal, not a single message — "add unit tests for the billing service", "wire up the new onboarding flow". Now you can hand the agent the whole thing.
What changed
- Plan · Build · Auto, right in the composer. Pick how much autonomy you want. Plan drafts an approach you approve first. Build is a normal chat turn. Auto runs the goal end to end.
- You stay in the loop. For a multi-step goal, the agent proposes a plan you can read and edit before any code is written, then works through it step by step.
- It checks its own work before declaring a step done (more on that in a separate note), and surfaces progress live so you can watch or walk away.
- It finishes with a pull request. Every change lands as a reviewed PR you merge — there are human approval gates throughout, and the agent never merges for you.
Why
The point of an agent fleet is leverage: describe the outcome, come back to a PR. The new goal modes make that the default path, while keeping the merge button firmly in your hands.