When something breaks
An assistant that works on its own breaks quietly. The key to your CRM expires, someone else's service stops responding, credits run out — and from the outside everything looks fine until somebody notices there are no replies.
That's why the platform writes to you itself.
Ask first, don't reconfigure
The shortest route is to tell the builder directly:
The bot isn't working
It looks first at what's broken, and only then changes anything. The most common cause isn't the settings but someone else's service — and reconfiguring something that works is the longest road to nowhere.
If there are no problems, it says so: everything the agent does outside is responding — and the cause has to be looked for in knowledge or character.
A complaint about content is a different question:
It gave a silly answer to the delivery question
Here the builder looks at the log of recent turns: which steps ran, which were skipped and why, which actions were called and what they returned. The cause will be named, not guessed.
Why the agent might have gone quiet
| What happened | What you'll see | What to do |
|---|---|---|
| no channels | “not answering anywhere yet” in the panel | connect a channel or put the chat window on your website |
| channel not set up | “channel not set up” in the panel | the channel exists but has no token — messages won't get through |
| out of credits | new conversations don't start | top up, or wait for the next period |
| spending cap | new replies are paused | the balance isn't the issue — you need a different plan |
| an action's key expired | the agent answers but can't run the action | issue a new key in your service and save it |
| someone else's service is down | same as above | nothing you can do but wait |
| the address leads to an internal network | the action, MCP server, page or link isn't called at all | it's not the service: the platform doesn't call local and internal addresses. Give a public address — Your own action |
| trigger auto-paused | the schedule no longer fires | fix the cause and turn the trigger on again |
| the turn runs out of time | the visitor sees an error instead of a reply | simplify the flow — it won't go away by itself |
| the agent “forgets” | asks again about things already agreed | the character isn't to blame — the conversation history wasn't saved |
| memory unavailable | greets a regular customer like a new one | a memory storage failure; the bot knows about it and doesn't pretend it's a first meeting, but will ask for details again |
| mailbox isn't being read | answers on the website and in messengers, but not emails | most often the password changed: tell the builder “my email password changed” and give the new one (for Gmail, iCloud, Yahoo — an app password) |
| the reply didn't get through | everything is green in the turn log, but the person didn't get a reply | check whether the channel is still connected and whether its token was revoked |
| not all the work was done | the summary says “all processed”, but only part was | the problem report says what exactly didn't get done: split the work into several runs |
| the task stopped | long work paused by itself: three passes in a row failed or did nothing, or the pass limit was reached | tell the agent or the builder “continue the task”; if the problem is the criteria — “change the task criteria”; if the work is no longer needed — “cancel the task”. What was collected stays — Long task |
| product catalog isn't updating | the store's product file can't be read; the assistant quotes prices from the last successful read | check whether the feed link opens; if the store gave you a new one — “here's the new link to the products”. Products stay, and the problem goes away by itself — Store products |
| the payment service rejects the key | the assistant can't give the shopper a payment link and honestly says paying won't work right now | the LiqPay keys were revoked or reissued: get new ones in your LiqPay account and tell the builder “update the payment keys” — Payment in the conversation |
| the payment service isn't responding | a payment link isn't created or the payment status isn't checked — including in the background check, where the problem appears after several failed attempts in a row; the shopper is told the truth, and an unchecked status is “unknown”, not “unpaid” | most often it goes away by itself; the problem closes as soon as a request goes through. If it persists — check your payment service account — Payment in the conversation |
| voice messages aren't recognized | the platform answers people who send voice messages with “please type your message”; the agent answers text as usual | this isn't in your hands: a failure of the recognition service or the messenger; the problem closes by itself as soon as a voice message is recognized. Don't need voice messages — “don't transcribe voice messages” (Voice messages) |
| images aren't being drawn | the agent tells people there won't be an image right now; posts and replies go out without images | not in your hands: the platform's image service isn't responding. The problem closes by itself as soon as the next image is drawn — Images and voice |
| text-to-speech isn't working | replies that should go as voice go as text; when asked to “read this aloud”, the agent says it can't right now | same thing: the platform's speech service; the problem closes by itself — Images and voice |
| calendar can't be read | when asked “are you free”, the assistant says the calendar couldn't be read, and doesn't book anyone | most often the secret address was reset or the app password revoked: “here's the new link to my calendar” or “update the calendar password”. The problem closes by itself — Calendar |
| notebook is full | the agent stopped remembering what it finds, while turns are “successful” | tell the builder “keep only the last 90 days in the set” (your number) or “throw away this set” if you don't need it |
Problems are visible in the dashboard too
Notifications go wherever they're set up to go — and if they're set up nowhere, there's still the dashboard. Open problems sit in the panel on the right above everything else, in the same words: the panel answers the question “what isn't working right now”. Fixed ones disappear from there.
Where notifications arrive
One cause — one message, however many times it repeats. Two hundred identical notifications overnight isn't care, it's a way to teach you to ignore them. When a problem goes away, the next message about it is “re-armed”.
There are three kinds of notifications, and you can combine them.
In Telegram — to the on-duty people in Telegram support. This is a separate bot where real people write, not the one your assistant answers in: notifications go to whoever is on duty in it right now. It works without setup as soon as support is connected and at least someone has sent it “/start”. The assistant's own bot doesn't send problem notifications while there's another place to send them (see below). You can turn it off on request:
Don't message me about problems in Telegram
By email. Often this is a shared support mailbox, not the owner's address:
Send problem reports to support@smak.example.com
Email isn't available in every installation
The emails are sent by the platform itself, and if it has no mail server set up, this option simply doesn't work. The builder knows this: it will say plainly that there's no email and name the options that are available — instead of asking for an address that nothing would ever reach anyway.
You can ask at any time:
Where will notifications reach me if the assistant breaks?
The answer covers this agent and this installation together: not “an email is saved” but “it will arrive by email” — or “it won't arrive by email, because the mail server isn't set up”.
To your system. The notification goes as a request to your address, so a problem can open a ticket right away:
Also send problems to https://hooks.example.com/youselfbot
None of the three — to your chat with the assistant. If you've linked your chat in Telegram, Viber, WhatsApp, Discord or Slack and notifications aren't set up anywhere, problems, visitor contacts and other alerts go there. As soon as at least one of the options above appears, they go that way — that's where the on-duty people and the team are, not just you.
What the agent itself writes to you (when its flow has a “notify the owner” step) goes the other way: to your linked chat, and only there; without a chat — as a notification. When you're talking to it directly (in the linked chat or in the “Speak as owner” window in the dashboard), there won't be a separate message: whatever the agent wanted to say, it says in its reply.
A failure on our side — the model is temporarily unavailable — also arrives here, the same way as problems, and no more than once every half hour. You don't need to do anything: this message just explains why the assistant is silent.
Customer requests and the digest are separate
A customer who asked for a person isn't a problem, and they reach you by other routes: your chat with the assistant, the on-duty people in Telegram and email — all at once, with one reminder and quiet at night. See Conversations → So requests reach you outside the dashboard. Once a week the platform also sends a digest: what the agent doesn't know and what's waiting on you.
What you'll be told
Not “an error occurred”, but what exactly to do:
The “create_client” action no longer has access: the key expired or was revoked. Issue a new one in the service it calls, and save it in the agent's keys.
The “morning summary” trigger has been paused: several runs in a row ended with an error.
The page “https://shop.example/teapot” can't be read — watching it has stopped: you won't hear about changes on it for now. Check whether it opens in a browser; if the address changed, tell the builder “watch the new address”.
On the website “smak.example.com”, which the agent took knowledge from, something changed that needs your decision. The agent no longer uses pages that disappeared from the website in its answers — but their texts haven't been deleted…
Credits are running out. When they're gone, the agent will stop — top up your balance in advance.
The monthly spending cap on replies has been reached. The balance isn't the issue: to keep it working, raise the cap by changing your plan.
You're warned about money in advance — when 80% of the monthly allowance is spent, that is, when a fifth is left.
The agent can react to problems itself
Platform events can wake the agent just like a schedule:
- an agent turn failed;
- a visitor asked for a real person;
- a visitor left a contact;
- credits are running out.
When a visitor leaves a contact, send it to me in Telegram right away
That way you don't have to wait until someone reads the notification.
Next
- If it didn't come out right — when the cause is your recent changes, not a problem.
- Can't sign in — if the issue isn't the agent but signing in.