Skip to content

Goal agent

The goal agent takes a goal in plain language, drives the device itself, and checks your acceptance criteria. A finished goal run can become a stored plan, which then replays with no model.

Before you start

  1. Set a model in Settings → LLM provider (Anthropic, OpenAI or LM Studio) and enter its key. Your provider bills you for every request.
  2. Start a live session from Setup → Start session. The goal agent needs one.

Run a goal

  1. Open Chat and turn on the Goal switch.
  2. Type the goal, for example: Log in as ${username} and add the first product to the cart.
  3. Click Run goal. The Inspector shows each action as the agent takes it.

Start from a ticket

In Sources, open a source and click Draft plan in chat. Chat loads the source and its acceptance criteria. The agent must verify each criterion before it can report a pass.

Always give criteria

A goal with no acceptance criteria verifies nothing. It can finish cleanly and still prove nothing about the app. Write criteria as observable facts: "The cart badge shows 1."

Keep the result

On a finished goal run card, click Promote to plan and name the plan. The plan keeps the platform and the environment of the goal run. Replay it from Plans like any other plan.