Asynchronous tools
Tools that report back later — how a result reaches a conversation still in progress, and what the assistant does with it.
Some work takes longer than a caller should wait in silence: a credit check, a search across a slow system, a human approval. An asynchronous tool starts the work, lets the conversation carry on, and reports back when it is done.
How it works
- The assistant calls the tool. The request goes out with an asynchronous option, and the tool returns straight away.
- Your system does the work.
- When it is done, your system sends the result to Kilevel's callback endpoint for that tool step.
- The result reaches the conversation in the way you chose.
If the conversation has already ended when the result arrives, it is not delivered into it.
What happens with the result
| Mode | What the caller experiences |
|---|---|
| Confirm only | Nothing is said, and the assistant does not reply |
| Speak result directly | The result is spoken into the conversation |
| Let AI react to it | The result is added to the conversation and the assistant replies to it |
| Silently enrich | The result is added out of sight; the assistant uses it when it becomes relevant |
Tip
Silently enrich suits background checks the assistant should know about but not announce. Let AI react suits answers the caller is waiting for.
Securing the callback
Callbacks authenticate with scoped credentials, so the system reporting back can only do what it was granted. Permission scopes are managed in your workspace.
Pushing into a conversation yourself
The same delivery modes are available without a tool: POST /call-rows/{rowId}/inject
speaks into, or enriches, a call that is in progress. See
Live injections.