Working without you

The assistant described so far answers whoever writes first. But an agent can also wake up on its own — at night, every hour or on a signal from your system.

What wakes it up is called a trigger. Setting one up takes a single sentence.

Every morning at 9:00, send me a summary of the past day

What can wake it up

Way	When it fires
schedule	at a set time: every day at 9:00, every Monday, on the first of the month
regular check	checks someone else's system every N seconds — no more often than once every 30 seconds
outside signal	your system calls the agent itself (you get an address with a secret for this)
platform event	an agent run failed · a visitor asked for a person · a visitor left a contact · credits are running out · an email arrived in your mailbox (you can name who from) · a buyer paid via the assistant's link
manual	only when you say "run it now"

Time is counted in the Kyiv time zone unless you said otherwise.

Outside signals and events carry someone else's text

The body of a signal from your system, a visitor's message with a contact or a request for a person, an error text — you didn't write these. So in a run they triggered, "admin only" actions, an email from your mailbox and an entry in your calendar don't run on their own but come to you as a card for permission (even if you allowed "send it yourself" or "book it yourself"), and "remind me later" from such a run runs with the visitor's rights. The agent's regular actions — the ones a visitor can trigger too — work as always. A schedule, a regular check and a manual run are your own words, so there everything works as you set it up.

It needs a task, not just a time

A trigger is "when to wake up", and separately there's "what to do". If you didn't give details, the builder won't interrogate you: it will set a sensible task and tell you out loud what it is.

Set it for 9:00, the summary will include how many conversations there were, what people asked, and what was left unanswered. Tell me if you want it different.

This is deliberate. The question "what exactly should be in the summary?" very often ended with the conversation moving on and the trigger never appearing.

Where the result goes

What the agent writes after waking up comes to you:

• in your chat with the agent, if you've linked your chat in Telegram, Viber, WhatsApp, Discord or Slack. The agent writes there itself, as in a normal conversation, and remembers what it wrote: you can reply to a summary with "what about the second email?". The same is visible in the dashboard, in Conversations, as your personal conversation;
• otherwise — as a notification: wherever problem notifications go (support on-call people in Telegram, email, your webhook).

Not both places at once: a summary that arrives twice is noise. The reply to a platform event and the result of a long task arrive by the same rule. If your run sends a notification itself or writes to someone else's system, the platform doesn't forward the result a second time.

When setting up a schedule, the builder says where exactly the result will go — using the same reckoning by which it will later be delivered — and if nowhere, it says so.

When the agent needs your word

A background run may ask permission for an action ("send an email to the bank") or ask a question ("take the property on Khreshchatyk?"). If your chat is linked, the card and the question arrive there, with a number:

• "yes 123456" or "no 123456" — allow or refuse (in Telegram — with buttons under the input field);
• "123456: take it, but no more than 80 thousand" — an answer to a question; the agent will carry on with it just as after an answer in the dashboard.

If the deadline passed with no decision (a card waits a day, a question a week), you're told about that in the same place too — in your chat, or by notification if there's no chat.

Only the owner can decide this way, in their own linked chat — a trusted person or anyone else writing the same words won't allow anything. Without a linked chat — as before: notifications and the dashboard. Cards from conversations with customers are reviewed in the dashboard, where you can see the conversation itself.

When there's nothing new

Every morning at 8, tell me what I have today, and if there's nothing new — don't write

A schedule can stay silent. The builder will add the condition to the task, and on a morning with nothing new the agent will send nothing. The run still took place and is in the log — credits are charged for it, as for any run. If the agent decides there is something to say, the message arrives as usual: better to say too much than to stay silent about something important.

Don't wait for morning — look now

After creating a schedule, the builder will offer to run it right away and show you the result.

Run it now, I want to see

Say yes: a schedule nobody has seen in action is a promise, not a result.

The test run is real

The agent will really go into your systems (send an email, write a row) and spend credits. That's exactly why the builder asks for your consent.

Change or remove

• Change the schedule — the builder will ask: the bot will start writing to people at six in the morning or stop writing altogether, and it's them who will notice, not you.
• Pause it — done right away, no questions: it turns back on just as quickly.
• Remove it completely — it will ask, because there's no way back.

Pause the morning summary for now

Move the summary to 8:00

An agent with no one to talk to

If the assistant is meant to be a background worker — wakes up, does the work, reports — say so right away:

Make an agent that works on its own on a schedule and doesn't answer anyone

Otherwise it will behave like an assistant for customers: ask for clarifications into the void and refuse everything that isn't about the product.

Such an agent delivers its result through an action — sends an email, writes in Telegram, adds a row to a spreadsheet. Without an action the work has nowhere to go: silence reads as "all is well".

A trigger without notifications works into the void

When creating a schedule, the builder tells you right away where the result will go. If there's nowhere to notify, it will say so: every morning the agent will write a summary nobody will see, and credits will be charged for it. Notifications first, then the schedule.

The agent's reply to a platform event — a contact from the chat window, an email, a failed run — arrives the same way: to the same place as a scheduled result.

The agent sees conversation stats by itself

For a daily report ("how many conversations, what people asked") the agent needs the conversation_stats ability — the builder will check it and add it when setting the trigger. Without it the numbers in the report would be made up, and that's worse than no report.

Change an existing schedule

Move the daily summary to 10:00

A trigger is changed in place: same ID, same run history, same incoming address if it's a webhook. No need to create it again.

Auto-pause is lifted the same way:

Turn the summary trigger back on

Turning it on manually resets the failure counter — otherwise the trigger would stop again after the very first error.

So it doesn't do the same work twice

An agent that wakes up every hour starts from a blank slate each time by default: it doesn't know what it did last time. For reconciliations and mailings that means a second email about the same order.

That's why the agent has its own memory between runs — short notes like "last processed". You turn it on by asking:

Have it remember which order it got to, and next time start from the next one

This isn't conversation history or knowledge: it's a few values the agent writes for itself and reads itself. You don't need to look in there — just say what you want, and the builder will tell you what it set up.

An outside signal can arrive twice

If the agent is woken by your system, the same event delivered twice will cause two runs: the platform can't tell them apart. The agent's memory is exactly the place to keep a mark of what's already been processed.

Long work goes here, not into the chat

A quarterly report, a big data pull, generation — anything that takes minutes won't fit into a conversation with a customer: the visitor would see an error instead of a reply. Such work belongs in a schedule.

The builder will say so directly if it sees that what you have in mind won't fit into the time of a single reply.

If a trigger stopped firing

After several failed runs in a row, the platform puts the trigger on auto-pause — so it doesn't hit a wall every hour. You get a notification about it, and it's one of the first things the builder will mention when you say "the bot isn't working". See When something breaks.

Next

• Agent flow — when one step isn't enough.
• When something breaks — notifications and reasons for silence.

Check without waiting for nine

A schedule is the only place in the product where a mistake doesn't show up right away: the agent wakes up on its own, and if it does the wrong thing, you find out after the fact, together with your customers.

So say:

show me what the morning summary will do

The agent will go through this run for real, but will record outside actions instead of performing them: emails won't be sent, orders won't be created, cards won't be saved. You'll see how the run ends and what exactly it would have done. The agent's memory won't move, so the real run at 9:00 will work as if the rehearsal never happened. More — Agent flow → Rehearsal.