Human Service
LoopAgent's human handoff is a capability triggered by the model's judgment: when the model decides a human is needed, it initiates the handoff. The core guarantee is that once a human takes over, the agent immediately stops answering.
Configuration
Go to the LoopAgent's "Human Service" settings module, which shares the same configuration as Agent / FlowAgent (the connected customer-service system, authentication, service hours, trigger timing, and so on). For the steps to connect and authenticate a third-party customer-service system, see Human Service.
- The authoritative switch is "Enable" in the human-service module. Turn it off and the human-handoff tool is not registered for that turn, so the model won't promise a human.
- Trigger timing can be described custom, helping the model judge in what scenarios to hand off.
The possible outcomes of a handoff
After the model initiates a handoff, it gets different outcomes, each with a different follow-up behavior:
| Outcome | Meaning | Follow-up |
|---|---|---|
| Success | A human conversation is created / transferred successfully | The model says a line and then stops |
| Outside service hours | Currently not within human-service hours | Sends your preset "outside service hours" message as-is to the user and ends the turn |
| Not ready / handoff failed | Not enabled, channel validation failed, all agents offline, etc. | The model relays the reason and keeps serving on its own |
Distinguish the two kinds of "didn't go through": outside service hours sends the preset message as-is and ends the turn; not ready / handoff failed has the model relay the reason and keep serving with AI.
The hard guardrail: the agent goes quiet after takeover
- Within the turn: after a successful handoff, any remaining tool calls in the turn are blocked, and the model can only say a single handover line.
- Across turns: while the human conversation is active, the agent is skipped entirely — user messages go into the human channel, and the AI no longer replying is expected behavior; it resumes automatically after the human conversation is closed.
Agent-facing note
During a handoff, the model generates on the spot an internal note visible only to the agent (the handoff reason, conversation summary, key facts, and suggested next steps), which the customer never sees.
- The note is at most 2000 characters; anything longer is truncated.
- Supported channels: LiveChat, LiveDesk
Troubleshooting when a handoff won't work
- Whether the human-service module is enabled;
- Whether it is currently within service hours (outside hours it directly sends the outside-service-hours message);
- Whether the channel passes authentication;
- For LiveDesk, whether an agent is online (all offline causes a handoff failure);
- Handing off within a customer-service system's own Widget is an "in-channel transfer," which requires both that channel's transfer switch and the integration-success flag to be on.