Agent Chat

Not a chatbot — describe a job and the agent works out what it needs, then does it. Also the single place every channel conversation arrives.

It acts, it doesn't just answer

Most chat boxes give you text back. This one does the work.

You describe what you want in plain language. The agent decides which capabilities the job needs — a browser, a connected app, a skill — assembles them, carries out the steps, and gives you the result. You do not have to know which pieces it will use, or name them.

Tip — If you find yourself writing a prompt the way you would write one for a chatbot, ask for more. "Summarise this spreadsheet" works, but so does "pull last quarter's figures out of this spreadsheet, chart them, and email the chart to my accountant."

A worked example

Here is a single prompt, exactly as you would type it:

Search on https://josefseibelusa.com/ find mens dress shoe black US 9 size,
take a screenshot and email it to my Gmail along with price and any discounts
available. Provide the purchase link also in the email summary. Save the email
summary and screenshot as a PDF and give me a link to download.

Nothing in that prompt names a tool. The agent works out that it needs to browse a website it has never seen, capture what it finds, send mail from your account, and produce a document — then does each part in order.

What comes back:

• An email in your Gmail with the price, any discount, and a link to buy
• A screenshot of the product page as it looked at that moment
• A PDF of the summary and screenshot, with a download link

Behind that, four separate capabilities were used — browser control, a screenshot tool, your connected Gmail, and a document skill. You named none of them.

Naming things yourself, if you want to

You can be specific when it matters. Alongside your prompt you can add chips:

Chips are an option, not a requirement. If you would rather not think about which toolkit or server applies, describe what you need and let the agent work it out. Both routes reach the same place.

Tip — Chips are worth using when the same job could go two ways and you care which — two mail accounts connected, or a result that must land in a particular channel.

When something is missing

If a job needs a capability your workspace has not connected, the agent says which one and stops rather than improvising a worse answer.

The fix is usually to connect it and ask again — a toolkit to authorise, a skill to install, an API key to provide. See Secrets, MCP servers and toolkits for where each of those lives.

Everything from every channel, in one place

Agent Chat is also where your channel conversations collect.

A message you send from Slack, and the agent's reply, appear here. So does a WhatsApp thread, a Telegram exchange, an SMS. It works the other way too — you can send from here and the message goes out through the channel it belongs to.

That means you do not have to remember where a conversation happened. Ask something over WhatsApp on the way to work, and it is in Agent Chat when you sit down.

See Connecting a channel for how each channel is set up.

Attachments

You can attach files to a prompt — a spreadsheet to work from, a document to read, a screenshot of something you want changed. Put your question in the same message as the file.

When the agent needs your approval

Some steps wait for you before going ahead: publishing something, spending money, sending a message on your behalf, or anything a Human in the loop chip has marked.

The agent pauses and asks in the same conversation. Nothing happens until you answer — a paused job holds its place rather than timing out into proceeding.

What it costs

Agent Chat spends billing tokens like any other agent work, and an elaborate job spends more than a question. The example above browses a site, captures an image, sends mail and builds a document — four pieces of work, not one.

If your balance is too low to run a request, the agent tells you rather than failing quietly. See How billing tokens work.

Tip — If you find yourself typing the same elaborate prompt repeatedly, turn it into a workflow. A workflow runs the same steps for fewer tokens and can be scheduled. See Workflows and skills.

Next steps

Agent Chat is for a job you want done now. For a job you repeat, continue to Workflows and skills — and for a whole function you want run, see the agent library.