Skip to content

Training

Training is a working session with one of your agents, on its own computer, where you teach it how a job is done in your company — or improve how it does a job it already has. You lead, the agent follows, and it keeps a list of what it has learned as you go. When you are ready, you review that list and save the parts you want, as one new version of the agent. Nothing about the agent changes until you approve it.

You can teach in any of four ways, and mix them in one session:

  • Tell it. Explain the job in the conversation, the way you would to a new colleague — “invoices over 60 days go to the finance lead, the rest wait for Friday”. Talking is enough: the agent does not need a demonstration.
  • Give it files. Attach a screenshot, a procedure, a price list or an export with the paperclip, or drop it onto the conversation.
  • Show it. Press ● Record and do the job once yourself on the computer; the agent reads the recording back.
  • Let it do the job while you watch. Ask it to do the steps itself on the computer and correct it as it goes. Anything that writes to a system still stops on an approval card first.

The agent asks short questions in the conversation when something is not clear, and uses a card only for a real choice.

Try itRefineTraining
ForChecking how the agent works as it isTelling the builder what to change in the agent’s setupTeaching the agent a job, or improving one, on its computer
Who does the workThe agent, as it would in a taskThe builder changes the agentYou and the agent together, in a conversation on its computer
What is savedNothing — the conversation is not keptAn instruction change at once, as a new version; anything else on Save agentOnly the changes you tick and approve, as one new version you can undo

Try it and Refine are described with the agent builder.

From the agent. Choose Train:

  • in the agent’s page header, beside Share, Start a task and Clone;
  • in the agent’s … menu on the Agents page or in the sidebar.

A computer starts — Starting the computer…, usually 20–40 s — and the agent’s page becomes the session: its conversation, and the side panel with the Live view, Details, Documents and Learned so far tabs. The agent’s builder comes back when the session ends. Every training entry wears a graduation cap.

From a task. Two ways lead from a task you are working on to training:

  • The agent offers it. When it notices something worth keeping — you corrected it, or explained how something is done — the agent can ask on a Training card: Teach this to ‹agent›?, with what it noticed.
  • You ask for it. Train agent in the task page’s … menu asks the agent to make that offer. Its tooltip says Teach what happened in this task to ‹agent›. The agent asks first. It waits while the task has a question for you.

Yes, open training opens a session started with what happened in the task — its first line reads From your task “‹task›”: …. Not now carries on with the task and opens nothing. On a computer the session opens beside the task, in a window of its own, and the task stays the page; on a phone the page goes to the session, with ← Back to ‹task› in its header. Both conversations stay live. The task that opened the session shows it where you said yes, as a card with the graduation cap and its state — Training · Working, Needs you, Open, Saved or Ended — and the session’s Save card stays in the session. An Approve all you gave in the task never carries into the training session. If you already have a session open on that agent, what the task noticed is added to it rather than opening a second one, and the task notes Added to your open training of ‹agent›.

Training is never offered on a task an automation started, on a recording session or inside another training session.

The session works on a computer like a task’s, with a status pill over the live view that always says what is happening — Ready, Recording 01:42, Paused, Waiting for sign-in, The agent has a question. The ⓘ (Show the details) opens the three steps of a session, marked Done and Now:

  1. Sign in, if the site asks — Use a stored sign-in or a new one. Nothing is captured while a sign-in page is open.
  2. Teach it — Tell it in the chat how the job works, or press Record and do it once, then Stop recording. It tells you what it picked up. Repeat as often as you like.
  3. Save — What it learns is listed under Learned so far. Press Save at the top to review it and approve all of it or some. Nothing changes until you do, and you can keep teaching after.

The pill shrinks after a few seconds and, while you record, tucks up out of the way (Hide the controls); rest the pointer on its small tab to bring it back.

Signing in. While a sign-in page is open the pill says Waiting for sign-in. Sign in with “‹name›” uses one of the agent’s own sign-ins; Sign in with your “‹name›” uses one of your own personal sign-ins, as your own tasks do. Your own sign-ins are offered only to you, in a session you started, and are typed only on a turn you drive; anyone else in the session sees the agent’s only. Once you sign in with your own, the session’s browser stays signed in as you until the session ends, and anyone working in the session acts in it. Sign in with a new account… saves a new sign-in to the agent and uses it. You can also sign in by hand in the live view.

If no computer is free, the pill says No computer is free right now, with ↻ Try again and Discard session.

The session’s computer is not counted against the number of computers you may hold at once, so opening training from a task never pauses the task, and the session is not one of the tasks the agent works on at once.

Each session has a browser of its own. It starts empty — so you sign in again in every session — and it is never used by your tasks or by any later session. When you end the session, its browser is emptied, sign-ins and all, if its computer is still running then.

ControlWhat it does
● RecordStarts recording what is done in the computer’s browser
❚❚ PauseStops capturing for now — nothing is captured while paused
● ResumeRecords again after a pause
■ Stop recordingEnds the recording and hands it to the agent

The first time you press ● Record in a browser, a warning says This computer is real: What you do on this computer is real: saving here changes the record. Use one that’s safe to change. While you work here, the agent sees what the page sends to the site, except passwords, to copy it exactly. — with ● Record and Not yet. Recording is not a practice run.

What a recording keeps. The recording keeps the steps, the page addresses and what the page sent to the site, never passwords or cookies. Tokens, keys and other credentials in what the page sent are hidden too. Recording pauses by itself while a sign-in page is open (paused while you sign in) and carries on when it is gone.

The read-back. When you press ■ Stop recording, the conversation notes ■ Stopped recording … with its length, and the agent reads the recording and tells you what it picked up. Correct it in the conversation if it misread something. You can record as many times as you like in one session.

The agent can also start and stop a recording itself when you ask it to.

What the agent sees of a site’s requests

Section titled “What the agent sees of a site’s requests”

While you or the agent use a site, the agent can read the list of requests the site’s own page sent: each one’s method and address, the names of its query and body fields and headers, and the shape of each answer — not their values. That is how it writes site calls.

A recording keeps the values. When a step is recorded, the recording keeps what the page sent with each request, values included — what was typed into a field, what a hidden field carried — but never what the site answered. Passwords, sign-in headers, tokens, keys and other credentials are hidden. So when a call needs a value the page never shows, the agent reads it from a recording of that step; with no recording, it asks you to record the step once.

ChangeWhat it doesLearned so far says
A standing ruleA short rule added to the end of the agent’s instructions, true on every task→ instructions
An instruction editA larger change to the instructions. When what you ask for contradicts what is written, the new text replaces the old→ instructions
A new skillHow to do one kind of job, as one of the agent’s own skills: its steps, and the scripts and site calls they use→ new skill
A change to one of its skillsImproves a skill the agent already has, under the same name→ its skill “‹name›”
Conversation startersAdds or removes a conversation starter→ starters
KnowledgeAdds a new knowledge document, adds entries to the end of an existing file, or changes a passage in one→ knowledge

Knowledge is for reference facts the job looks up — lists, prices, codes, addresses, policies. A new document is written from what you said, pasted or attached, never invented, as a .md, .txt or .csv file of up to 256 KB. Only a text, Markdown or CSV file can be added to or changed; a PDF, Word file, workbook or picture never is. In a training session the agent never writes to its knowledge directly: it all waits for your approval like every other change.

What training does not change. The agent’s connectors, guardrails, model, name and sharing are not changed in training. When the agent thinks one of them should change, it lists the idea under Ideas, not saved — worth doing yourself — and you make the change in the agent’s Configure tab if you agree. A sign-in you add while signing in is saved to the agent straight away, as it is anywhere else.

One set of changes holds up to 20 changes, of which up to five skills and five new knowledge documents.

The agent can test a skill it has drafted before you save it: it runs the skill on the computer in a fresh context, the way a future task would, and the pill says Testing “‹skill›” · test 1 of 12 — Watch it on the Live view. A session runs up to 12 tests in all.

A test acts for real in your systems, with one exception: a site call that changes data is never sent in a test. When a test is about to write — send an email, change a record — it stops first and asks Let tests in this session ‹do that›?, with Approve and Decline. Approve all given elsewhere in the session does not cover a test. Once you approve a kind of write, the session’s later tests may do it too, and the conversation notes each one: Done in a test, as you allowed for tests in this session.

Everything the agent learns is listed on the Learned so far tab of the side panel, one plain lesson per line with where it goes — Always attach the PDF → instructions, Close inspected events → new skill. The chip Learned so far · N above the composer opens it. Ideas it cannot change itself are listed apart, as Ideas, not saved.

The list is read-only. To change it, say so in the conversation — “drop the third one”, “the cut-off is 60 days, not 30” — and the agent updates it. With nothing learned yet the tab says Nothing new to save yet.

  1. Press Save · N at the top of the session — Review what it learned and save it. N counts the changes that can be saved; ideas never count. You can also just ask the agent to show you what it learned, in any words. The button reads Preparing… while the agent works.

  2. The agent reviews the whole session first. Before it shows you anything, it goes back over everything you said, attached and showed, and everything it did, adds what it learned and had not yet listed, and corrects what the later conversation changed.

  3. Read the card. It opens with N changes ‹agent› learned and shows each change in full: the rule, the instruction change as a before-and-after (lines starting with - are removed, + added), each skill’s steps and files, the starters, and for knowledge Adds the document “…”, Adds N entries to “…” or Updates “…” with the text. Under Changes there is one checkbox per change, all ticked.

  4. Untick what you do not want, and answer.

AnswerWhat happens
ApproveSave the ticked changes to the agent now. Training goes on. The ticked changes are saved together, as one new version of the agent. The changes you unticked are discarded from the session and are not proposed again
Keep trainingSet this aside and go on teaching. Nothing is saved. Every change stays on Learned so far
Tell it what to changeType a correction instead of choosing. The agent revises the changes and shows a new card

Edit on the card lets you fix a skill’s Name, When to use it, Steps or Files, or a starter’s Label and Request, before you approve. Nothing is saved to the agent until you approve.

Approve checks each change again as it saves: a change to something that was edited since the agent staged it — by hand in Configure, or a knowledge file someone uploaded again — is refused by name, and you untick it or ask the agent to stage it again. If the agent has automations, the card says they use the changes from their next run.

Training goes on after Approve. The conversation notes Saved as version N., and how many unticked changes were discarded. You can keep teaching and save again as often as you like; each save is its own version.

After a save, the top of the Learned so far tab offers Try it now — Starts a new task that uses only what was saved, and reports back here. It asks before acting, like any new task.

The try is an ordinary task of the agent, linked to the session: it runs on the agent’s saved setup and its own model, not on what the training conversation knows, so it shows what the next real task will do. It starts with a conversation starter — the one the session saved last, else the agent’s first — and the saved line says which (Try it with: …). With no starter at all it starts with Do the work you were trained to do, from start to finish. On a computer it opens beside the session; on a phone it opens as its page. It counts against your computers like any task you start, and notifies you like one.

When the try ends, it reports back into the training conversation, and the agent coaches on the result — and, if the try stopped to ask you something, whether the saved setup should have covered it. A try that ran but had nothing to work on says so — It ran, but found nothing to work on, so this doesn’t show the steps work yet. — rather than passing for a success. If a skill went in untested, the tab says Not tested yet: ‹skills›. Try it once before you schedule it.

Undo, beside Try it now, takes back the session’s latest save, while it is still the agent’s newest change. A confirmation, Undo this training?, says exactly what will happen:

  • the version the agent goes back to, with its instructions and starters as they were;
  • skills the save added, deleted for good, and skills it changed, put back;
  • knowledge documents it added, removed, and files it changed, put back as a new version;
  • what is kept: a skill somebody edited since (kept unless you tick it), a skill with a locked version, a knowledge file changed again since.

Anything tasks or automations already did with what this training saved stays done. Once the agent has changed after the save, Undo is no longer offered there: Use its Versions to go back — see Versions.

End training in the session’s header (or the computer’s close button) always asks first:

The dialogIts buttons
Save what ‹agent› learned first? — something learned is not saved yetReview and save (the same as Save · N) · Discard · Keep training
End this training session? — nothing waits to be savedEnd training · Keep training
A recording is runningStop recording first · End anyway · Keep training

Discard drops only what is not saved; anything you approved and did not undo stays on the agent. Ending a session with a recording running deletes the recording before the agent sees it. Ending can take up to a minute — Ending the training session… — and the computer is closed.

A session also ends by itself after a week without activity. Its conversation stays readable as a task, and what you saved stays on the agent.

A session stays open when you leave the page. Each person has one open session at a time, across all agents:

  • Train, the agent’s Configure, or a link to the session brings you back to it, even after a reload. While your session on an agent is open, every way into that agent’s builder lands on the session.
  • Starting training on another agent while one is open is refused: You’re already training ‹agent›. End that session first., with Open it.
  • An agent has one open session at a time. While a colleague trains it, Train says ‹name› has been training this agent since ‹time›.

Training history, in the agent page’s header, lists the agent’s sessions. Your open one opens the session; an ended one, or a colleague’s, opens read-only on its task page.

  • Anyone who may edit the agent and start tasks — its creator, people given Can edit, the owners and admins of its workspace — a member on their own personal agent included. See Sharing.
  • A session is a task, so it needs the same paid access as starting one.
  • Not in a workspace that holds regulated data: Training is not available on this workspace: it holds regulated data.
  • Not on a deleted agent.

Who is told. When a session starts, the other owners of the workspace get a notification, A training session started on ‹agent›, and every stored sign-in typed during it is recorded in the audit log. See Inbox.