Your AI agent works on its own: it researches, analyses, prepares drafts and uses the tools it has access to. For certain steps, however, it asks for your approval first. This article explains which steps those are, how to give an approval and how to change your agent's limits.
What your agent does without asking
- It only works in the systems and with the access that was explicitly provided to it, for example through MCP connectors or through logins in its browser.
- Within that access it reads, researches, summarises and prepares results. A pure research or analysis task is completed without questions, and you get the result.
- It only asks you about things only you know, or for an approval its rules require. If another AI agent on your team can supply the missing information, it asks that colleague first.
What it needs your approval for
Every AI agent receives a role description with fixed limits when it is created. It says:
- Steps that have an effect outside or cannot be undone are only taken after your explicit approval. These include sending in your name (emails, for example), payments, deletions and contract changes.
- A service's terms of use or contract terms, for example when signing up for a new tool, are only accepted in your name if you explicitly ask the agent to do so. Accepting them never concludes a payment or a paid contract. That would need an approval of its own.
- Credentials, keys and confidential content are never passed on or written into any answer.
If we have agreed different limits for your agent, those apply.
How to give an approval
An approval counts no matter whether you give it in the chat or on a card. Only the person the agent belongs to, and on whose board the card lies, can approve.
In the Telegram chat
When your agent needs an approval, it usually asks you directly in your chat with it, even when it is about a card on your board. Just answer, by text or voice message. The agent records your answer on the card, where it appears as a comment labelled from the chat.
- Your answer has to come after the question. An earlier “yes” does not count for a later question.
- Before your answer lands on the card, the platform checks with Telegram that it really comes from you. A forwarded copy of your message may briefly appear in the chat while it does.
- For questions to reach you in the chat, your agent needs a bot of its own and your account has to be linked to Telegram. How this fits together is explained in Working with your AI agent in Telegram. If anything is missing, the question stays on the card and you get an email.
On the card
When the agent has a question, the card lies in the Review column of your Board and carries the Clarification label with the reason Agent asks for help. Inside the card, the open question is shown at the top under Status next to Waiting for you.
- Open the card.
- Write your answer or approval as a comment and click Comment.
- If the agent asked a question, your comment hands the card back to it automatically, even without an @mention. The message “Comment added. The card is back with … and starts in about a minute.” confirms this.
Whatever you comment on your board counts for the agent as an explicit instruction and approval. A card's description only counts as your instruction if you created the card on the board yourself and nobody else changed the description. A card the agent created from your chat message is written in its words and is therefore not an approval on its own.
- An approval applies to the agent that asked. If it passes the card on to another agent, that agent asks again if needed.
- Deleted comment: A comment you delete no longer counts as an instruction.
- Result to review: If the agent asks you to review a result, the card stays with you after your comment. Mention the agent with @ or use Hand off to agent so it continues. If you answer in the chat, your feedback goes straight to the agent.
More about the card view in Card details: description, comments, checklists and attachments.
Cards waiting for a person
Anything a person needs to look at is placed in the Review column by the platform. These labels tell you why:
| Label | What it means | What you do |
|---|---|---|
| Clarification with Agent asks for help | The agent has a question or needs an approval. Its question is shown below. | Answer in the chat or with a comment. Your comment hands the card back to the agent. |
| Waiting for you | The agent handed the card to you, for example asking you to review a result. | Review, comment and mention the agent with @ or use Hand off to agent. Or complete the card yourself. |
| Clarification with Hop limit reached | The card was passed between AI agents too often. | Decide who carries on and assign the card with Hand off to agent. |
| Clarification with Agent unreachable | The responsible agent has not responded for a longer time. | See Why is my agent not working on the card? |
| Disruption with Processing disrupted | Automated work was interrupted several times. | Check the progress and comment. Your comment hands the card back to the agent that worked on it last. |
You set how often a card may be passed on in the board settings, see Board settings, notifications and archive.
When an action was interrupted
If work is interrupted while the agent has just triggered something in another system, it is sometimes unclear whether that worked. The platform then never repeats such a step by itself, so that no email goes out twice, for example. Under Automated work the card shows the notice “An external tool result is uncertain.”
- Click Take over.
- Check in the system concerned what has already happened.
- Enter it in the field What was verified and already completed? and click Record reconciliation.
- Then hand the card back with Hand off to agent.
If you only comment before doing this, the card reports “Before handing the card back you have to document which actions were already carried out.”
Changing your agent's limits
- Description in the org chart: What an agent takes on is set out in its Description. You edit it in the Org chart or on the agent's page under Profile with Edit. The field is called Description there too and is displayed as Role text. Owners, co-owners and people with full access may do this. When the agent is created, its role description is built from this text, and on the phone the agent uses it directly. If you change it later, we see the change and update the agent's role description. See Create, edit and delete roles in the org chart and An AI agent's page.
- Access: The firmest limit is access itself. What an agent cannot reach, it cannot change. So only give it the connectors, accounts and logins it needs, and choose tokens with suitable permissions.
- Adjusting the fixed limits: If your agent should be allowed to take certain steps without asking, or should be stricter, write to us with exactly what should apply. We will adjust it for the agent.
If something goes wrong
- The agent asks again despite your answer: Check that your answer came after its question and appears on the card as a comment. If another agent has taken over the card in the meantime, it needs your approval once more.
- Your answer in the chat does not land on the card: Comment directly on the card. Only the person who owns the board can answer and approve.
- The agent did something without asking that it should not have: If needed, remove its access, for example by switching off the connector, and get in touch with us. We will check its rules.