# How to work with the agent

The agent works as a doer, not a search box: it pulls data from your connected services and files, does the work and hands back a result — a spreadsheet, a report, a letter, a message.

This guide is about phrasing tasks so you get what you need on the first try.

## The first 10 minutes

### Connect a service you use every day
Go to **Integrations**: marketplace, CRM, email, bank, telephony, accounting — wherever you normally export numbers by hand.

### Ask for a report you already make yourself
Not "what can you do", but "assemble what I assemble every Monday". You will see the result on your own data straight away.

<TryPrompt prompt="Build a sales report for last week: revenue and margin per product, plus the 10 biggest drops. Deliver it as an Excel file, one row per product." />

### Look at what comes back
The agent replies in the chat and attaches a file — spreadsheet, document, deck — that you can forward as is.

### If you need it regularly, say so
One sentence turns the report into a [scheduled task](/help/tasks) that arrives on its own.

**About the question “what can you do”**
It is the most common first question and the least useful one: the answer is a list of capabilities that leads nowhere. One real task of yours says more about the product than any description.

## What to know right away

- **New topic — new chat.** The agent re-reads the whole conversation at every step, so a sales report and a supplier contract do not belong in one chat. See [Chats and context](/help/chats).
- **Send exports as files.** Spreadsheets, documents, PDFs, photos: a file lands in the working environment and the agent computes over every row. A short fragment is fine to paste. See [Files and results](/help/files).
- **Name the service with `@`.** A connection alone does not tell the agent which account this task needs. See [Connected services](/help/integrations).
- **Make changes in small groups.** Edits go straight into your service, so "all the cards" is worth replacing with a condition and a limit. See [Changes in your services](/help/changes).
- **Standing rules go to memory.** Say "remember: …" and the agent will apply it in new chats. See [Agent memory](/help/memory).
- **Repeating work becomes a task.** If you are copying the same request a second time, put it on a schedule. See [Scheduled tasks](/help/tasks).
- **A site with no integration — through the browser on your computer.** A supplier portal, a government service, an internal admin panel: the agent opens them on your screen. See [Browser on your computer](/help/local-browser).

## Cheat sheet

1. New topic — new chat.
2. In the request: what to take, what to do, in what form to deliver, for what period.
3. Exports as files, short fragments as text.
4. Service via `@`.
5. Standing rules as "remember: …".
6. Anything regular as "create a task: … every Monday at 10:00".
7. Corrections should be pointed, not "redo everything"; changes in a service start as a list, then a small group.
8. Make your first task a real one, not "what can you do".

Next: [how to phrase a task](/help/prompts) so you do not have to rephrase it.