# What a task is

> A task is one thing you ask Ask agent to do. Its stages, Plan first, and why a deploy waits for a person.

Prompt for your AI: Read https://mousemouse.ai/docs/tasks.md first. Explain how a task of Ask agent moves from a plan to a deployed policy, and where I decide. Ask me before anything that costs money or moves a robot.

A task is one thing you ask Ask agent to do, in your own words. "Train
Laika to walk forward at 0.3 m/s" is a task. So is "Test this policy in sim".
The agent plans the work, runs it, and shows you the results. Every task
appears in one list on the [Ask agent](/agent) page.

Ask agent appears where this hub has an AI model connected. Elsewhere, connect
your own AI to the hub instead: see [Use the hub from your AI](/docs/mcp).

## Starting a task

Type what you want and press **Start task**. The chips under the box add
context: the robot, a dataset, a policy. **Runs on** says where training and
sim tests run: Lucen cloud CPU, Lucen cloud GPU, or your own machine. Lucen
cloud appears only where this hub offers it. The agent must keep to the
choice you make.

The examples under the box name only robots and policies you can see. A click
fills the box. Nothing starts until you press **Start task**.

## Plan first

**Plan first** is on unless you switch it off. With it on, the agent shows its
plan and waits for you. The plan lists each training round, where it runs,
and what it may cost. You approve it, or tell the agent what to change.

With Plan first off, the hub may approve a plan for you. It does so only when
the plan is free, or fits what is left of your free allowance. Any other plan
still waits for you. Only you can switch it off, from your own signed-in session.

A plan that would go past your free allowance needs one more step. You tick
the box that accepts the extra amount, and only then can you approve it.

## The stages

A task's page shows which stage it has reached, and when it reached each one.

1. **Planning**. The agent is working out what to do.
2. **Awaiting approval**. A plan, or a spend the plan did not cover, waits for
   you.
3. **Training**. A training run is going.
4. **Testing in sim**. A sim test of the policy is going.
5. **Ready for review**. There is something for you to read. It may be a
   finished run and its sim test score, a shortlist, or the agent's answer.
6. **Deploy pending approval**. A request to run the policy on a robot waits
   for a person.
7. **Deployed**. A robot runs the policy, or ran it.

A task can also end **Failed** or **Stopped**. Sometimes a decision waits for
you while work goes on. The task then keeps its stage, and the list adds
"Needs you".

The **Evidence** section of a task's page plays the newest sim test video. It
lists every sim test of the task with its result, and every deploy request.

## Deploying needs a person

The agent can train a policy and test it in sim. It can also ask to run the
policy on a robot. Only a person signed in to the hub can approve that. The
request waits on its own approval page, with the checks the hub ran.
[How approvals keep your robot safe](/docs/approvals) explains each step.

You can also give a task to one of your connected robots. Choose it in the
Robot chip. The agent then works on that robot, and a person approves each
time it hands the robot a policy.

## Stopping a task

Press **Stop** on the task's page. The agent stops, and so do the training runs
it started. A robot already running a policy keeps going until its approved
minutes end. To stop the robot now, press **STOP** on [My robots](/settings/devices).