Call handling
The call-lifecycle limits on an agent - maximum call duration, silent callers, and hanging up on an abusive caller.
Call handling decides how long a call may run and when the agent should end it. Every control here is enforced on live calls, and every workspace sees the same controls.
Where these settings live
Everything here is stored on the agent, not on a version. Reverting to an earlier version does not restore old call-handling values.
A published version is still read-only, so create a draft to change these controls.
Maximum call duration
A hard ceiling on how long a single inbound call may run, between 60 seconds and 14400 seconds (4 hours).
Leaving the field blank does not mean no cap: a blank value applies the platform default of 15 minutes. A real inbound call is never uncapped. The setting can only raise or lower the ceiling, never remove it.
Not a sudden hang-up. Three things happen in order.
- A wind-down. Shortly before the cap, the runtime tells the agent to start wrapping up, and the agent speaks that in its own words on its next turn.
- A wait for a seam. At the cap the runtime waits briefly for a natural gap rather than cutting across a sentence. It also stands aside for a warm transfer already in flight.
- The closing line. The agent speaks one short line and the call ends.
The wind-down lead is the smaller of 40 seconds or a quarter of the cap.
15 minute cap winds down at 14 minutes 20 seconds
1 minute cap winds down at 45 seconds, not 20 seconds inThe closing line
Spoken by the runtime, not the model, so the caller always hears why the call ended even if the model is mid-thought. The default:
Sorry, we've reached the time limit for this call, so I'll have to let you go now.
Override it with Time limit closing line, up to 300 characters. Keep it short because it takes the floor over whatever was being said.
What the cap covers
Inbound phone calls only. In-app test calls and outbound legs such as onboarding and campaign calls are not covered by this setting.
Silent callers
Two controls under Silent callers handle a caller who goes quiet.
| Control | Default | Range |
|---|---|---|
| "Are you still there?" check-in | On | On or off |
| Silence before the check-in | 15 seconds | 5 to 120 seconds |
With the check-in on, the agent asks once after the silence threshold:
Sorry, are you still there?
If the line is still quiet after a second silence it signs off and ends the call:
No worries, I'll let you go now, take care.
Both lines are spoken directly, without a round trip to the model. The silence count resets the moment the caller speaks.
Turning the check-in off
Leaves the line open without prompting. An idle call still hits the platform's own idle backstop.
Abusive callers
Hang up on abusive callers is on by default. It works in two halves.
The prompt half
A short protocol is added to the agent's instructions, drawing a sharp line between a frustrated caller and a bad-faith one. Swearing about a problem, being curt, repeating themselves and pushing hard on price are explicitly not abuse.
Only a caller who stays abusive after the agent has kept its composure earns a calm warning:
I'm not able to continue like this. Is there something I can help you with?
If they ease off, the agent carries on normally. Only if they persist may it close with "Thank you for calling. I'll leave it there." and end the call.
The agent warns at most twice, and it does not argue, lecture, or accuse.
The runtime half
The end-call tool refuses an abuse hang-up unless the agent actually spoke the warning earlier in the call. Only the agent's own turns count, so a caller repeating the line back cannot arm the hang-up.
When the protocol is added at all
The setting must be on and the agent must have an end_call tool. An agent
that cannot hang up gets no protocol. Renaming the tool does not break this:
the platform matches on the built-in, not the name you gave it.
Turn the setting off for lines where ending a call is never the right answer: superannuation, health, government and crisis services.