Skip to main content
When you’re chatting with Swan directly, some actions pause and wait for you to confirm before Swan carries them out. This keeps anything destructive, or anything that fans out into a lot of work, in your hands. Writes started from an app do not pause. They run immediately, and a tool set to always ask does not stop them. Swan still rejects data tools and tools that require alert-level approval.

What asks for confirmation

By default, Swan asks before it:
  • Deletes or archives something. Deleting a Campaign or a skill inside Swan, or archiving a record (company, contact, deal, or task) in a connected CRM like HubSpot, all require your confirmation first. This is irreversible, so Swan won’t do it silently. Approving a Campaign deletion that doesn’t name a destination Campaign archives the Campaign’s sequences, which stops every active one. Deleting a sequence and excluding a company run without confirmation by default.
  • Removes something from a third-party tool. Deleting a lead, account, opportunity, or note in Salesforce, deleting a record in Attio, deleting a sequence or lead in outreach tools like Instantly, or removing a contact from a sequence in Reply.io or Lemlist all ask first. Everyday creates and updates in those same tools (adding a contact, updating a field) do not require confirmation by default — only the delete/remove actions do.
  • Delegates more than 6 tasks at once. When Swan hands off more than 6 tasks in one call to run independently (for example, “research each of these 40 accounts”), it asks you to confirm before dispatching. The confirmation shows that this will trigger multiple independent runs and may consume significant credits, so you can decide whether to go ahead or narrow it down. Up to 6 tasks start without confirmation. Once you approve, a table of the individual sub-agent runs appears inline in that same chat conversation so you can track each one’s progress.
  • Submits a form on a website. When Swan browses the web and has typed into a page, the next click or key press that could send the form asks first, showing the site and the button or key involved. Browsing and reading pages never asks, and this check can’t be changed from Tools.
Everyday actions — searching, reading data, drafting a message, editing a sequence, changing inbox or outreach settings, updating your own org’s records inside Swan — don’t require confirmation.

Changing what requires confirmation

An admin can override the default for most individual tools from the tool’s entry in Tools — open the toolkit’s capabilities list and set a specific tool to always ask for confirmation, or to never ask, regardless of the built-in default. The same list can disable a tool entirely or restrict it to Admins. This applies per organization, not per user.

Deciding in the app, in Slack, or on WhatsApp

If a run is happening in a chat conversation and needs your confirmation, you can approve or reject it either from the chat itself (the confirmation appears inline in the conversation at agent.getswan.com/chat) or, if the conversation is happening in a Slack thread, from the Approve/Reject buttons there. Whichever channel you decide in first is the one that counts — once a decision has been recorded, the other channel’s buttons no longer apply to that action. On WhatsApp, Swan lists the pending actions and asks you to reply APPROVE or REJECT. The decision is all-or-nothing: it applies to every pending action in that conversation at once. It also has to be typed — a voice note saying “approve” is transcribed like any other message but is never treated as a decision, so a misheard recording can’t authorize an action. Sending a new image or file while Swan is waiting for approval rejects all pending actions and starts a new request in the same conversation, preserving the attachment and caption. This rule is shared by all chat channels that accept attachments. Rejecting an action doesn’t stop the run. Swan treats the rejection as new information: it asks you why, or proposes a different way to accomplish the same goal, and keeps going from there. In Slack and WhatsApp, sending a new message while Swan is working interrupts the active request, like a force send in the app. Swan waits for the old run to stop before handling the new message in the same conversation. If it cannot stop the old run, it asks you to retry instead of running both requests at once. Approval waits still follow the rules above.

Trigger and background runs never pause

Runs that happen automatically — a trigger firing on a schedule, a webhook, or another background event — never stop and wait for a person the way a chat conversation does. If Swan hits an action that would normally require confirmation, it skips that specific action instead of executing it, and asks a human to approve or handle it separately (a Slack message or a task on your Desk), then continues the rest of its work. The action itself is never carried out without a human eventually approving it — the run just doesn’t sit there paused. See how triggers work and AI behavior and guardrails for more on this distinction.

Common questions

Why is my chat input locked or read-only? This usually means there’s a pending confirmation waiting on an action, or Swan has delegated work to sub-agents that are still running in the background. Resolve the pending approval, wait for the sub-agents to finish, or press Stop (the send button in the chat input turns into Stop while a run is in progress) — stopping cancels the run and any sub-agents it dispatched, and the input unlocks. Slack, WhatsApp, and MCP conversations automatically continue through long-running work in the background. Long chat turns can also pause briefly while Swan keeps working, and when delegated sub-agents finish. With the conversation open, the answer keeps streaming. If you close or reload the tab, reopen the conversation within 3 minutes to keep watching it, or press Stop to cancel the request. After 3 minutes Swan finishes the request in the background, and the answer is waiting when you return. If I approve in Slack, does the app update too? Yes — approvals and rejections are recorded against the run itself, not the channel you used, so the other surface reflects the same decision. Can I require approval for every single tool Swan uses? An admin can set individual tools to always require confirmation, but there’s no single switch that makes every action in every toolkit require approval. Does rejecting an action cancel the whole task Swan was doing? No. Swan asks what you’d like it to do differently and continues working from there — it doesn’t abandon the run. Why didn’t Swan ask me before updating a HubSpot contact? Routine creates and updates to CRM records don’t require confirmation by default — only deleting/archiving records and similar destructive actions do, unless an admin has changed that tool’s setting.