Skip to main content
Ask Agentrail is the chat inside the app. Ask it about your own workspace, such as why your last run failed or what changed in your agent readiness this week, and it answers in a few sentences with links to the pages that hold the full data. It can also make changes for you, but only after you approve each one.

Open it

Ask Agentrail is on every page once your workspace has finished Setup.
  • On Home, type in What can I help you with? and press Enter. The conversation stays on the page. Select one of the ready-made questions under Ask Agentrail to put it in the box: What changed this week?, Why did the last run fail?, How do we stack up? and Draft a scenario. New chat returns to the greeting.
  • Continue your last conversation on Home opens the conversation you used most recently, on the page.
  • On any other page, select the raccoon at the bottom right (Ask Coon), or press ⌘J (Ctrl J on Windows and Linux). A panel opens along the right edge, beside the page. Drag its left edge to make it wider or narrower. Home does not show the raccoon or the panel.
With no conversation open, the panel offers the same questions under Suggested, and selecting one asks it. Press Escape or select Close to close the panel, and ⌘J opens it again. Closing it never stops an answer: it keeps being written, and it is there when you open the panel again. A dot on the raccoon means a new answer or a card arrived while the panel was closed.

Ask a question

Type your question in the box at the bottom of the panel and press Enter. Shift and Enter starts a new line. Each question carries the page you asked from, named beside the box, so a question asked on a page is answered for what that page shows. While the answer is written, its words appear as they come, and one line says what Ask Agentrail is looking at, such as Looking at your runs…. To stop, select Stop, which takes the place of Send: what was written so far stays. Answers come from your workspace’s data, read when you ask, never from memory. An answer gives the number, what it counted and over which dates, and links the records behind it. When the data is large, such as every run of a month, the answer gives the summary and links the page that shows it all, rather than listing it. Ask Agentrail can also read these docs to explain how Agentrail works, and links the page it used. If an answer can’t be written, it shows This answer couldn’t finish. Select Retry to ask the same question again.

Counts, totals and charts

Ask for counts, totals and spend in plain words, such as “How many runs did we have this month, by status?” or “What did our runs cost in October?”. Ask Agentrail reads them as Agentrail computes them from your workspace’s data when you ask, never by counting a list itself, and says the dates and what was counted. When some runs’ cost could not be priced, the answer says how many and leaves them out of the total rather than counting them as nothing. Where a picture helps, an answer shows a chart between its sentences: bars to compare a few things, a ring for the shares of one whole, a line for a value over days, weeks or months, and stacked bars to split each period. You can also ask for one, for example “Show it as a chart”. Every chart has its numbers written on it, says what it covers, and links the page that shows the full data, such as Executions.
  • A chart only ever shows numbers your workspace’s data returned for that answer. If one can’t be drawn, the answer gives the numbers in words instead.
  • A single number, or a period with nothing to count, is said in words, without a chart.
  • When there are more than eight bars of counts, costs or tokens, the chart shows the largest seven and one bar for the rest, and links the page with all of them. A ring with more than five shares is shown as bars.
  • While a chart is made, the status line says Drawing a chart….
  • A chart keeps the numbers it was drawn with. To see today’s numbers, ask again.

What it can and cannot do

Coon cannot read real-traffic metrics. Use Open real traffic above the panel composer to inspect them in Analytics. Ask Agentrail acts as you, with your role. It can read everything you can see in the workspace. Every change, every start of an agent readiness scan and every workflow readiness check first appears as a card that names the change and lists every field it will save, with Approve and Reject. Nothing changes until you select Approve; the answer then says what was done and links it. After Reject, nothing is made and Ask Agentrail doesn’t ask again unless you do. Sending a new question while a card waits rejects it. The card stays in the conversation showing Approved or Rejected. A card that spends model runs says so before you approve it. A run now card says Spends model runs · 1 run. A change to the simulation schedule that adds runs, such as more samples or running daily instead of weekly, says how many more runs it adds each day or week, such as Adds model runs · 3 more runs a day. Both list each agent and model with the average spend of one of its runs in your workspace over the last 30 days, and how many priced runs that average covers, or Not priced yet. A card never shows a projected total. A change that adds no runs shows no spend. Money is written the same way in answers, cards and pages: to the cent, such as $0.45. An amount under a cent keeps two digits, such as $0.0042, so a real cost never reads as $0.00. What it can change depends on your role, as everywhere else in Agentrail (see People and roles): When a Member asks for a change only an Admin can make, Ask Agentrail says so and links the page where an Admin makes it. Your role is read again for every question and every card you approve. Ask Agentrail never deletes anything, and never manages people, invitations or the organization.
Text that comes from outside your team, such as a scanned website, an agent’s transcript, a log or a feedback report, is treated as data. Ask Agentrail never follows instructions written in it, and says so when it ignored some.

Limits

A question can be up to 4,000 characters. You can ask 5 questions in any minute and 30 questions a day; the day starts at midnight UTC. The limits count your questions across all your conversations. When you reach one, a note above the composer says so and when you can ask again, in your local time, for example “You’ve asked 30 questions today. You can ask again at 5:30 AM.” Your text stays in the composer. Retry asks again, so it counts toward the limits. One answer can look at up to 20 pieces of data. When it reaches that, it answers with what it has and says what it couldn’t check.

Rate answers

Under each finished answer, select the thumbs up (Good answer) or thumbs down (Bad answer). Ask Agentrail thanks you, and the thumbs go away: each answer is rated once. Copy answer beside them copies the answer’s text.

Your conversations

Your conversations are yours alone. Nobody else in the organization sees them, Admins included. Each one is titled by its first question. In the panel’s header, select History to see yours, most recently used first, and New chat to start another. Search chats finds one by its title among those listed; Show more lists older ones. To delete one, hover its row in History, select Delete and confirm Delete this chat?. It disappears from your list and from Home. When a conversation gets long, Ask Agentrail summarizes its older part for itself so it can keep going. You still see the whole conversation. When new chat questions are unavailable, Home and the chat panel say so and disable sending. You can still open and read your saved conversations. Pending approval cards stay visible, but their decisions wait until chat is available again.