# Plan and progress

> On long jobs the agent shows its live plan, delivers each milestone in the thread and leaves the final answer below it.

URL: https://www.ghosty.studio/en/docs/plan-and-progress

When you hand it something with several steps —"review the six contracts and open a PR for each fix"—, the [Teams](/en/docs/channels/teams) agent works the way a teammate would in a thread: first it says what it will do, then it keeps delivering, and at the end it sums up. You don't have to ask "how's it going?".

## What you see in the thread

1. **The plan, on top.** The agent's first bubble is its task list, and it is **edited in place** while it works:

   | Mark | Means |
   |---|---|
   | ✓ | done |
   | ✱ | in progress (shows what it is doing right now) |
   | ○ | pending |

   The footer says when it last changed. Below the plan are the steps and the tools it used.
2. **Milestones, as messages.** Each deliverable is posted as soon as it exists —a finding, an opened PR, a finished file— as its own message in the thread. It doesn't wait for the end.
3. **The answer, at the end and separate.** When it finishes, the result is posted as a new message **below** the plan, so you can read it without scrolling past the list.
4. **It mentions you only when it needs you.** An @mention from the agent means something needs your action (approve, decide), and it notifies you even mid-job. Everything else doesn't.
5. **It corrects what it already said.** If something it posted turned out wrong, it edits that same message: it strikes out what no longer holds and adds **[Edit: …]**. The message is marked as *edited*.
6. **A footer on each answer.** Below it says which agent answered and with which model; whoever manages the workspace also sees **Configure**, which opens the agent in Studio.

If you ask for something else in the same thread, the new plan takes the main spot and the previous ones collapse to **"Latest task list →"**, which jumps to the current one.

A short request ("what does clause 4 say?") has no plan: it answers in a single bubble, as always.

## Jobs that take hours

A job can last hours: the agent keeps going while there is work, even if a command takes long (a build, a long test run).

If the platform restarts mid-job, a job **with a plan** resumes on its own within seconds: the agent gets its plan and what it had already done, and continues with what's pending in the same thread. A short request that gets cut leaves the **Resume** button.

## How to ask for it

There is nothing to switch on. It helps to say it in the request:

- "Do it step by step and tell me every time you have something."
- "Post each PR as soon as you open it."
- "When you're done give me the count: how many, what's left."

While it works you can **correct** or **stop** it by writing in the same thread: the message enters the running turn.

## Which engines do it

| Engine | Live plan | Milestones in the thread |
|---|---|---|
| Claude | yes | yes |
| Codex | yes | yes |
| Others | no (they answer in one bubble) | depends on the engine |

The engine is chosen when you create or edit the agent in Studio.