How do you route tasks dynamically to different agents based on intent or confidence?

What you need to know

Three routing styles

Legacy 0.2 used a custom speaker_selection_method callable and allowed_or_disallowed_speaker_transitions.

Code first, model second

1def route(messages):2    last = messages[-1]3    if last.source != "user":4        return None                          # let the selector model decide5    text = last.to_text().lower()6    if any(w in text for w in ("refund", "chargeback", "money back")):7        return "billing"8    if any(w in text for w in ("router", "no internet", "slow speed")):9        return "tech"10    return None                              # unclear: model picks1112team = SelectorGroupChat([billing, tech, general], model_client=client,13                         selector_func=route, termination_condition=stop)

Handoffs with Swarm

1from autogen_agentchat.teams import Swarm2from autogen_agentchat.conditions import HandoffTermination, MaxMessageTermination34triage = AssistantAgent("triage", model_client=client,5    handoffs=["billing", "tech", "user"],6    system_message="Work out the issue. Hand off to billing or tech. "7                   "If unsure, hand off to user with one clarifying question.")8team = Swarm([triage, billing, tech],9             termination_condition=HandoffTermination(target="user")10                                   | MaxMessageTermination(15))

Each entry in handoffs becomes a tool like transfer_to_billing. When the model calls it, a HandoffMessage moves control. Handing off to "user" stops the run so your app can ask the customer and resume.

Confidence that means something

A model saying "confidence: 0.9" is not calibrated. Better signals:

• A small intent classifier's probability, and the gap between the top two intents.
• Whether required fields were found (order id, phone number).
• Retrieval scores for the question against each specialist's knowledge base.

High confidence goes straight to the specialist. Low confidence or a narrow gap goes to a generalist that asks one clarifying question, or to a human.

A real-life example

An e-commerce customer-support triage team handled 20,000 chats a day with a plain SelectorGroupChat. The selector call added about 700 ms and ₹0.05 per turn, and 8% of chats went to the wrong specialist first, for example "my refund for the broken phone" went to tech because it mentioned a phone.

They added a keyword selector_func and a small classifier on the first message. Chats where the classifier was above 0.85 and ahead of the runner-up by 0.3 routed directly (72% of traffic). The rest used the LLM selector, and messages with no order id went to general to ask for one. Mis-routes fell to 2.5%, and they added a hop cap: after three handoffs, the chat goes to a human.

Follow-up questions to expect

• "Selector or Swarm?" — Selector when a coordinator should see the whole picture; Swarm when each specialist knows its own limits and the next owner.
• "How do you debug mis-routing?" — Log each selection with the candidates and the last message; most fixes are better description text or a new rule in selector_func.
• "What stops handoff loops?" — A hop cap, a MaxMessageTermination, and a fallback to a human after N transfers.