Long task
"Find all land plots in Madrid of a hectare or more" isn't one reply. It's several hundred sites, each of which has to be opened and read, and at least an hour of work.
That's what a long task is for. You tell the agent what to find, it replies "got it" — and keeps working on its own, pass after pass, until it's done. All this time the task is visible in the dashboard: how many sources have been checked, how many results found, where it is now.
| Schedule | Postponed work | Long task | |
|---|---|---|---|
| when it wakes up | every time, by the clock | once, at a named time | pass after pass, until it's done |
| who creates it | you | the agent while working | the agent, when asked for something big |
| how it ends | it doesn't | with one run | with the word "done" and a summary |
How to set it up
In plain words — to the one who will be searching. For an agent the builder has taught to run long work, it's enough to write:
Find land plots for sale in Madrid from 10,000 m². I need price, area, location and a link.
If something's missing, it asks one question and starts. If the agent can't do this yet, tell the builder:
let this agent search the internet for anything I ask and build a database — it may take a long time
The builder will give it the abilities it needs and ask for your consent: a long task spends credits on its own, pass after pass.
How it works
A pass is a few minutes of work. In each pass the agent does the next batch: finds sources, opens pages, puts what it found into its notebook. Between passes there's a half-minute pause, and so on to the end.
Between passes the agent remembers exactly two things: the cards it has already saved, and a note to itself — "went through forty portals, a hundred and twenty left". So the work doesn't start over and isn't lost if a pass breaks off.
It keeps a status on each card: a source is "checked" or "not checked", a result "matches" or "doesn't match". Progress is counted from these statuses.
What it doesn't do
- It doesn't bypass site restrictions. If a site forbids automated reading of a page (the
robots.txtfile), the agent doesn't read it and marks the source with a separate status — it doesn't count as checked. You can open such a page yourself. - It doesn't go where a login is needed, unless you've given it access to that account.
- It doesn't visit one site more often than the site asks: there's a pause between pages of the same site.
Sources that couldn't be read are shown in the progress as a separate number — so "checked 347" doesn't mean "41 of them didn't open".
Where to see progress
In the panel next to the conversation there's a "Task N" row: how many tasks aren't finished yet. Click it and "What the agent is working on" opens:
- status — working now, running, paused, finished;
- progress in numbers — for each set: how many cards and how many in each status. The platform counts this from what the agent saved; it's not the agent's own retelling;
- note — where it stopped and what it will do next;
- task cards — everything this particular task found.
The list updates itself while the task runs. You can also just ask the agent: "how's the search going?" — it will answer with the same numbers.
Stop, continue, change the criteria
| Say or press | What happens |
|---|---|
| Pause / "stop the search" | the pass already running finishes, there won't be a next one; what's collected stays |
| Continue / "continue" | a pass starts right away |
| "search from 5,000 m², not from 10,000" | the criteria are replaced; the next pass reviews what's already collected against the new ones and keeps searching by them |
| Cancel / "cancel the task" | final: you can't continue, only set a new one. What's collected stays in the notebook |
Cancelling asks for confirmation — if you just want to hold the work, there's pause.
A dashboard admin can manage a task: each of these actions decides what the agent will spend credits on. Any member can see tasks and cards.
The result as a spreadsheet
When the agent finishes, it writes a summary wherever the task was set. The results themselves are in the notebook, and you can take them as a spreadsheet: the Excel and CSV buttons are both in the task cards and in everything collected.
The file gets all the cards of what you're looking at, not just the page shown. The columns come from the card fields: key, status, then everything the agent recorded — price, area, location.
The source base grows by itself
There's one card in the notebook, however many tasks come across it. If the first task found five hundred real-estate portals in Madrid, the second one doesn't search for them again: it takes the known sources and starts from them. Each task still has its own status — what the first one "checked" is still "not checked" for a second one with different criteria.
So searching gets faster over time: the agent starts not from scratch but from what it already knows.
When a task stops on its own
Every pass costs credits, and nobody is watching it. So the platform pauses the task itself — and says why:
| What happened | Why it paused |
|---|---|
| three passes in a row ended with an error | something broke, and repeating further would burn credits |
| three passes in a row with no progress | nothing changed in the notebook and the note is the same: there's nowhere else to search, or the agent is going in circles |
| three hundred passes without the word "done" | that's a day of non-stop work; from here you decide |
You'll learn about it from a "task stopped" problem, and whoever set the task — from a message in the conversation. What's collected stays. To carry on — "continue", and if the reason is the criteria — change them.
Limits
| Limit | How much |
|---|---|
| tasks of one agent at once | 3 |
| of them — not created by you (by a visitor in a conversation, by another program) | 2: one slot is always left for you |
| one pass | up to 5 minutes |
| pause between passes | half a minute |
| passes since the last "continue" | 300 |
| results per search results page | 8; the agent pages on by itself, up to the tenth page |
A task's cards are in the agent's notebook, so its limits apply to tasks too: fifty thousand cards per agent.
An agent's notebook is shared by everyone who writes to it. If different people use one agent, each sees only their own tasks in the conversation, but the collected cards sit side by side — for separate databases, create separate agents.
What to say
| Say | And it |
|---|---|
| "find all … by these criteria" | creates a task and starts the passes |
| "how's the search going?" | shows status and progress in numbers |
| "stop" / "continue" | pauses / continues |
| "search by different criteria: …" | changes the criteria and reviews what's collected |
| "cancel the task" | cancels for good (asks for consent) |
Next
- Deep research — a ready recipe on top of a task: branches search at the same time, the report is a document with sources.
- Search base — a search agent on top of a task: a source registry, results with filters, saved searches, publications.
- What the agent collected — cards, statuses and the spreadsheet.
- The panel alongside — the "Task" row among the others.
- Credits — how much a run costs and how to give an agent a monthly cap.