Squads

Squads let you break a complex workflow into specialized agents that hand off to each other. Each agent keeps a focused prompt, its own tools, and a clear job — instead of one mega-assistant with conflicting instructions.

Why use a squad?

Large all-in-one agents with long prompts and many tools tend to:

• Hallucinate more — the model loses focus across conflicting goals
• Cost more — longer prompts burn tokens on every turn
• Respond slower — large contexts increase latency

Squads split work at functional boundaries (qualify → book → confirm, or extract → enrich → notify). Handoffs keep the session alive while the next specialist takes over.

How it works

1. Create a squad and add member agents (each with its own tools and prompt).
2. Draw handoff edges on the canvas — when to transfer, what to say, what variables to extract.
3. Mark an entry agent — that member starts the call or process run.
4. Route traffic to the squad (voice channel / outbound) or invoke it via API (process).

Two kinds

• Conversation squads — canvas, Squad Console, channel routing
• Process squads — invoke, tools, Observe
• Handoffs & context — edges, variables, context policy
• Squad settings — share one agent's models or storage across the squad

Workflow framing

Treat a squad like a workflow graph:

• Nodes = agents (each owns tools, prompt, guardrails)
• Edges = handoffs (trigger description + optional variable extraction)
• Entry node = first speaker / first process step

Keep squads small. Split only when there is a clear boundary (lead qualification → booking, triage → specialist).

Dashboard

API automation

Programmatic management: Squads API. Process runs: POST /api/v1/invoke/. Outbound voice can pass squad_id on Calls.