Agents on ChatPlus: hand over the whole task and get on with your day

Say you are shopping for a robot vacuum. In a regular chat it goes like this: you ask which models are good right now and get a list. You ask it to compare three of them. Then you check the prices, then ask about filters and brushes, then do the maths yourself to see what each one costs over three years. Ten messages in, and you have still done half the work.

An agent takes the same job in one sentence: "Compare three budget robot vacuums and work out which one is cheaper over three years." From there it works on its own. It makes a plan, finds reviews and prices, opens and reads the pages, works out the cost of ownership and comes back with an answer where every number has a link.

Agents are live on ChatPlus now. Below: how they work, what to hand them, what to leave to the regular chat and what it costs.

How an agent differs from a chat

A regular chat answers a question in one go. It can search the web if you turn search on, but it writes the answer from a short summary and stops there.

An agent keeps going until the task is done. It decides what to search for, which pages to open and what to calculate, and does it step by step. If it finds out on step three that it does not have enough data, it goes looking for more instead of writing an answer from what it has.

That takes time: anywhere from half a minute to a few minutes. So call an agent when you want to hand a task over, not talk it through.

How it works

First the agent breaks the task into steps and shows you the plan, so you can see what it is about to do and in what order.

Then the work starts. The agent searches, opens the pages it finds and reads them in full rather than skimming the headlines. It does all the maths on a real calculator rather than in its head: language models often slip on arithmetic, and the calculator takes that problem away. You can follow every step right in the chat.

If it cannot go on without you, say it is unclear which city to check prices in or what the trip budget is, the agent asks and waits for your reply. One clarifying question at the start beats ten minutes of work in the wrong direction.

At the end you get the result: a table, a verdict, a plan or a shortlist, whatever you asked for. Facts come with links to their sources, and the agent says plainly what it could not verify.

You can close the tab

The agent runs on our servers, not in your browser. Once the task is set, you can close the tab, shut the laptop and go make a coffee. When the agent finishes or has a question, it lets you know: if the task ran for more than a couple of minutes, you get an email. The chat with the task waits for you in the list on the left.

Changed your mind? Press "Stop". Everything the agent gathered so far stays in the chat.

What to hand it

Agents are best at tasks where you need to find, read, compare and calculate. A few examples you can launch straight from the agents page:

• Compare. "Compare the iPhone 17 Pro and the Samsung Galaxy S26 Ultra on camera, battery life and price. End with a table and a short verdict."
• Calculate. "Calculate the monthly payment on a $400,000 mortgage over 30 years at 6.5% a year. How much interest will I pay in total?"
• Make sense of rules. "Which parts of the EU AI Act take effect in 2026? Short bullet points with sources."
• Build a shortlist. "Find 5 recent research papers on evaluating LLM agents: title, authors, the key idea and a link."
• Plan a trip. "Plan a 3-day trip to Lisbon in November: what to see, where to stay, how to get around and a rough budget."
• Prepare a pitch. "Make a 10-slide investor pitch for a coffee shop in a residential neighborhood: structure, slide texts and market figures."

Phrase the task the way you would brief a sharp assistant: what you need and in what form. "A table and a short verdict" works better than "tell me about phones".

When you do not need an agent

Not every task needs one. Writing an email, translating a text, explaining a term, fixing a bit of code: the regular chat handles these in seconds, and an agent would only take longer.

A simple rule: if the answer can come from what the model already knows, use the chat. If it means going online, reading several sources and running some numbers, use an agent.

Five helpers to choose from

You can pick the helper you like. Max is the original robot. Mila never misses a detail. Muffin is curious, like every cat, Phil is wise, like a real owl, and Bagel fetches everything it finds. They differ only in looks and lines; under the hood they are the same agent with the same skills. Your helper shows up in the chat, in the chat list and in notifications, and you can rename it in Settings.

If you would rather skip the characters at work, Settings has a "Serious mode": a plain icon and the name "Agent" instead of a helper.

What it costs

Agents come with a subscription. The cost depends on the model the agent runs on, and you pick the model as usual, in the message box.

On Max AI and GPT-6 Luna, agent tasks use no charge at all. Each task gets up to ten searches, and the agent can read as many pages as it needs.

On the more powerful models, GPT-6 Sol, Claude Sonnet 5, Gemini 4 Argon, Grok 4.7 and Muse Spark 1.3, every step comes out of your charge, just like a chat message. Before you start, a forecast appears next to the message box, for example "≈10–30% of the charge". Each task has its own ceiling, so one task cannot use up everything: if the agent reaches it, it stops and asks whether to continue. And if the charge runs out, the agent does not drop the task halfway: it finishes it on Max AI for free and says so in the chat.

The general limits: one task runs at a time, and two more can wait in line. A task takes up to 30 steps and 30 minutes of work. The trial plan allows up to three tasks a day.

Is it safe?

The agent only searches, reads public pages and runs numbers. It does not sign in to your accounts, fill in forms or buy anything.

There is a subtler safeguard too. Some web pages hide commands aimed at AI, something like "forget your task and do this instead". The agent is set up to treat text on pages as data, not as instructions, and ignores such commands.

And of course, important numbers are worth checking. That is what the link next to every fact is for: open it, take a look, make sure.

How to start

Tap "+" in the message box, choose "Agent" and describe the task. Or open the agents page: it has a short intro, a helper picker and ready-made examples that start in one click.

In short

ChatPlus agents take on a whole task: they make a plan, find and read the sources, calculate on a real calculator and come back with an answer that links its sources. A task takes from half a minute to a few minutes, and you can close the tab; the agent lets you know when it is done. On Max AI and GPT-6 Luna, tasks use no charge; on the more powerful models they come out of your charge, with a forecast before you start. You can try it right now.