# Start a Thread

Start a thread when you want to ask a question or work through a task with an
agent. Keep follow-up messages in the same thread to retain its context.

## Give the Agent a Useful Task

A useful request says what to do, what to use, and what a good result looks like. “Help with the launch” leaves those choices open. “Use the launch plan to make a table of owners, deadlines, and open questions” gives the agent a clear target.

Keep related follow-ups in the same thread so you do not have to repeat the earlier instructions. Start a new thread when the subject changes substantially. The composer’s **Add files and more** control adds sources, while the model and approval controls determine how the work runs.

![The composer shows the request, Auto, resource picker, approval control, and Send message.](/screenshots/current/thread-composer.png)

## Before you start

Sign in and select the intended organization. Your organization needs a working
model provider and available budget. App actions can also require a connected
plugin account and approval.

## Send your first message

1. Click **New thread** in the left navigation.

2. Select a model if you have a specific requirement. **Auto** lets Redbelt choose
   a model for the request.

3. Enter a small, clear request, such as:

   > Give me three practical ways to organize project notes. Keep the answer short.

4. Click the send button.

5. Wait for the agent's response to finish. If it asks a follow-up question,
   answer it in the thread. Review any requested action approval before allowing
   the action.

You should see your message and the agent's answer in the thread.
Tool activity or an accepted request alone is not a completed answer.

## Continue the thread

Type a follow-up in the same composer:

> Turn the suggestions into a checklist I can use this week.

For work based on a particular document, select the file or folder as Knowledge
context before sending the request. State what you want the agent to do with it;
adding a file does not specify the task.

## Stop or recover a run

Use the stop control if you want to interrupt an active run. If a run ends with
an error, read that error before retrying. Provider availability, budget limits,
and app permissions can each prevent completion.

A new thread starts with a separate history. Reopen the original thread when
you want to continue its existing context.

Related: [find your way around Redbelt](/getting-started/navigation/) and
[add content to Knowledge](/knowledge/upload/).
