Activity

Activity is every conversation one agent has had, one row each, newest first. It is the second page of every agent (a Team Agent has none), and it lists only that agent’s conversations; Insights › Activity is the same list across all of your agents. A row is called a run: it carries when it started, the first thing the person asked (the first 140 characters of it, never a later turn), which agent took it, how many messages it ran to, and how it ended.

Insights tells you the shape of a month. Activity is the list underneath it — the place you go when a number looks wrong and you want to read the conversations it came from.

How a run ends#

Every run carries exactly one outcome, and it is worked out from what the conversation did rather than stored on it. The order below is the tie-break: the first line that matches wins.

So a conversation that both handed over and captured a lead is counted once, as handed over. The counts in the key above the list are the whole window’s — worked out once, over the same runs and the same outcomes the list is drawn from, and never added up a second time from the rows on screen or from a separate count that could disagree with them. So the key holds still while you search, narrow by state or origin, or press Show 20 more: it is telling you what the window holds, not what the list under it is showing at that moment. A key that re-counted whatever was left after a filter would move under you every time you typed, and you could no longer compare it with the number you started from.

Filtering, and the window#

The strip above the list holds a search box (it matches the question text), and four controls:

• State — all states, or one of the four above.
• Origin — the site the conversation came from, for agents embedded in more than one place.
• Window — Last 24 hours, Last 7 days or Last 30 days. Thirty days is what you get on arrival.
• Sort — Newest, or Most messages.

If nothing in the window matches, the list says so and offers Clear filters, which empties the search box and sets State and Origin back to all. The window stays where you put it.

Twenty runs at a time#

The list shows twenty runs, with a Show 20 more button under them when there are more. The button says how many are left, and it never promises more than it can give you — with eighteen left it reads Show 18 more (18 left). When nothing is left it goes, rather than sitting there saying zero.

Changing anything in the strip above — the search box, State, Origin, Window or Sort — puts you back on the first twenty of the new list. Clearing the search box counts, so a list you come back to starts at the top as well. That is deliberate: the rows you had opened up are not the rows a new filter matches, so keeping the count would leave you scrolled into the middle of a list you have not seen the top of.

The twenty is about what is drawn, never about what was read. How far back the read itself went is a separate limit, and when it bites the page says so on its own line above the list — Showing the most recent … — so Show more is only ever a promise about rows that are already here.

One agent, or every agent#

An agent’s own Activity shows that agent’s runs and nothing else, and its counts are worked out from those runs. Insights › Activity is about your account, so runs from all of your agents sit in the same list. When the window spans more than one of them, each row names the agent that took it — on an account with a single agent that word would be the same on every row, so it is left off. Under the title is a line saying which of your agents the runs actually span, out of how many you have — a combined number without its denominator is a number you can’t check.

Nothing is ever shown as “running”#

There is no live or in-progress state. Nothing is written down while an answer is streaming — the conversation is stored once it exists, so a run appears when there is something real to show rather than flickering through a status nobody can act on.

Expanding a row#

Clicking a row opens it where it sits and shows its steps — what the agent did in that conversation, in order, with the moment it said it didn’t know marked as one of them.

Teach it sits on a declined step. It opens that agent’s own knowledge, on the What it knows tab, with the run’s opening question already in the Correction box — the same first 140 characters the row shows, which is not necessarily the turn that was declined. You write the answer and save. That agent’s is the load-bearing word: Activity covers your whole account, so the row you expanded may belong to any of your agents, and the correction is filed against the one that declined rather than whichever agent you happened to have open.

The review screen below has its own way to correct a wrong answer, and it does a slightly different thing — see the next section.

Opening one#

Open on any row opens the review for that conversation: the session details, the whole transcript, the outcome, and What it used — the sources the answers cited. The chip at the top says how the conversation resolved — handed over, a lead, or answered. On a wide screen it opens beside the list; on a narrow one it takes the list’s place, and Back to Activity brings the list back.

Try this again in Test sends that conversation’s first question into the agent’s Test dock, so you can see what it says now, after you have changed something.

Two controls sit on that screen, and they are the same ones you get arriving from People or from Insights:

• Fix this answer — every assistant turn in the transcript carries its own button, not just the run’s first question. Opens a short pop-over on that one turn: choose Give this answer, Say it doesn’t know (prefilled with your own words from When it can’t answer, if you set them), or Hand off, save, and it tells you whether the fix is live now or waiting for your next train. This is the wider one: the Teach it on a declined step, above, only appears where an answer was refused and always scopes to the run’s opening question — Fix this answer reaches a wrong answer anywhere in the transcript too.
• Delete this conversation — removes the exchange and its messages. What it leaves behind is deliberate: a follow-up record or a gap that came out of the conversation stays, so your outcome counts don’t silently change. See Where a lead came from.

Replay steps through it one stored step at a time; see Replay a conversation.

What Activity doesn’t show yet#

These are marked where they would have gone rather than left blank or filled with a zero. The first three sit on the run itself, under the steps, once you expand a row:

• Tool timings — how long each step of an answer took. Timings aren’t stored yet.
• Searches run — which searches an answer made. Also not stored yet.
• Actions before today — an action run before 2026-09-02 carries nothing tying it to the conversation that triggered it. Every new one does.
• Summary — a one-paragraph rollup of a conversation, on the review screen. Nothing generates one today, so the card says so rather than showing a paragraph nobody wrote.

Search across the console is on the same list: the Search row in the sidebar is a marker, not a working command palette yet. (The ⌘K search on these docs pages is a different thing, and it does work.)