What the agent collected

An agent that goes through hundreds of sources every day has to remember not one marker but a set: a thousand found properties with price and link, a list of platforms with publication status, a month of a post's metrics history.

These are cards — and they are not knowledge. The difference is simple:

	Knowledge	Cards
what it is	what the agent uses to answer customers	what the agent acts on
where from	you gave it: a site, text, a file	the agent found it itself through its work
what it looks like	a topic it retells in its own words	a row with a key, a status and fields

The main thing: it won't do the same thing twice

Every card has a natural key — a link to a listing, a phone number, a post ID. The platform itself tells the agent whether it has seen this key before. That's why "don't show me the same thing twice" works with no setup at all: the agent puts what it found into its notebook, and a repeat simply updates the card instead of adding a second one.

Without this, every daily pass would report "340 new found", a number that means nothing.

Where to see them

In the panel next to the conversation there's a "Collected N" row. Click it and "What the agent collected through its work" opens: sets with numbers, cards inside.

There you can:

• filter by status — show only "in moderation" or "rejected";
• find by word — search across card fields;
• find by meaning — in the conversation with the builder: "find something like 'income property by the sea'" will find a card that says "apartments, seafront, ROI 7%", even though they share no words. It searches this way only where meaning-based search is turned on on the platform; if not, the builder will say so and offer to search by words — it doesn't pass off similar letters as similar meaning;
• correct by hand — a wrong price or phone number is fixed in place;
• throw out — a dead source will no longer take up the agent's attention;
• take it as a spreadsheet — the Excel and CSV buttons: the file gets all the cards in the set, not just the page shown.

The platform knows the "sources", "results", "contacts" and "publications" sets more closely: it brings them to a common look and shows them in a separate section — Search base.

Next to each set there's a breakdown by status — "520 cards — checked 347 · not checked 132 · unavailable 41". That's the progress of the work, counted by the platform.

Correcting and throwing out are admin-only: a card is data the agent will act on tomorrow.

You name the sets

"Properties", "platforms", "investors", "drafts" — the names are yours. The platform doesn't invent its own and doesn't know any columns in advance: a card holds exactly what you asked it to hold. This isn't a CRM and doesn't try to be one.

So a set doesn't grow out of control

There are limits, and it's better to know them in advance:

Limit	How much	What happens when you hit it
cards per agent	50,000	new ones aren't saved; the agent stops remembering what it finds
sets per agent	32	a new set isn't created
size of one card	16 KiB	the card isn't saved; it's no longer a row but a page — such things belong in knowledge

A publication log for two hundred platforms hits the ceiling in about seven months. So say it in advance:

keep only the last 90 days in the "platforms" set

The agent will remove older cards itself. Age is counted from the last touch, not from when it appeared: a card the agent saw yesterday isn't old, however many months ago it first appeared.

If a set was created by mistake — "throw out the set 'day-2026-05-01'". That's the only way out when you've hit the set ceiling.

When the ceiling is actually reached, you'll learn about it from a "notebook full" problem, not from silence: the run doesn't fail — it finishes successfully and reports work that wasn't saved.

Cards as a work queue

The most useful thing here isn't the list but working in batches. Say:

go through all the platforms that aren't published yet

The agent will take a batch of unprocessed ones, do the work on each and change the card's status. The next run will take the next batch — no cursor needed, because the card's status is the cursor. You see the progress in the panel instead of guessing how far it got. More — Agent flow.

Cards of a long task

When the agent runs a long task, the cards stay the same — one per key — but each task has its own status. A source the first task marked "checked" is still "not checked" for a second task with different criteria.

That's why a task has its own entry — "Task cards": only what it found, with its statuses, and a spreadsheet of just those cards. The "Collected" row shows the whole notebook — the base that grows from task to task.

What to say

Say	And it
"keep a list of what you find"	sets up a notebook with deduplication by key
"show me what you collected this week"	shows the sets and cards
"find something like 'income property by the sea'"	searches by meaning, not by words — if meaning-based search is on on the platform; otherwise it says so
"fix the phone number in this card"	changes the field
"keep only the last 90 days"	removes older ones itself
"throw out this set"	removes the whole set (asks for consent)

Next

• Long task — work bigger than one reply: progress, pause, changing criteria.
• Agent flow — batch processing, rehearsal and the limits of one run.
• Working without you — the schedule by which the agent fills its notebook.
• The panel alongside — the other rows that appear when there's something to say.