Multi-Agent Voice Handoffs: How CallSphere Routes 37 Agents

According to Metrigy's CX Optimization 2025-26 study, 37.6% of companies plan to fully replace IVRs with AI triage agents. Among research-success-group companies, that number jumps to 62.5%. CallSphere runs 37 specialist agents across 6 verticals on this pattern.

What changed

The handoff pattern replaced press-1-for-sales IVRs in 2026. Three reasons it works now:

1. Triage models got cheap. Claude Sonnet 4.6 ($3/$15) and GPT-5 mini classify intent in ~150ms. Pre-2025 the triage call alone cost more than the conversation it routed.
2. Handoffs got first-class. OpenAI Agents SDK, LangGraph, and CrewAI all ship handoff primitives. Glue code is no longer required.
3. Voice latency is now sub-second. Streaming TTS plus efficient handoff context transfer keeps the turn-time under the human-perception threshold (around 800ms).

The architectural pattern: Triage agent classifies intent and hands off to one of N specialists. Specialists own their domain. Specialists can re-trigger triage if intent shifts mid-conversation. Specialists can escalate to a human when uncertainty crosses a threshold.

Why it matters for production agent teams

Three production wins from the handoff pattern:

Specialist quality beats generalist quality. A real-estate buyer intent specialist with 6 tools and a 3-page prompt outperforms a 30-tool, 12-page jack-of-all-trades. Tool-call accuracy goes up; latency goes down.

Per-vertical iteration. When the mortgage flow needs a new tool, you ship it to the Mortgage agent only. No regression risk to the Property Search agent.

Per-vertical cost control. Cheap intents (FAQ, status check) route to a Haiku-class specialist. Expensive intents (mortgage qualification) route to an Opus-class specialist. Cost-per-call drops 40-60% vs running everything on the heaviest model.

The Metrigy data point that matters: companies in the research-success-group (those measuring AI ROI) are 62.5% planning full IVR replacement vs 37.6% for the broader market. The pattern is winning where teams measure it.

How CallSphere applies this

CallSphere's production inventory: 37 agents · 90+ tools · 115+ DB tables · 6 verticals · 57+ languages · HIPAA + SOC 2.

Three concrete deployments:

Real Estate OneRoof: 10 specialist agents on hierarchical handoffs.

Triage -> Property Search -> Suburb Intelligence -> Mortgage -> Compliance -> Booking

Triage uses Sonnet 4.6 ($3/$15). Specialists use a mix of Sonnet 4.6 and Opus 4.7 for the hardest reasoning steps (mortgage qualification, compliance review).

IT Helpdesk U Rack IT: 10 specialists with ChromaDB RAG.

Triage -> L1 Diagnostics -> L2 Hardware / Network / Auth Specialists -> L3 Engineering Escalation

Each specialist has a focused RAG corpus (network specialist sees only network KB articles).

After-hours overflow: 7 agents organized as Primary + Secondary + 6-fallback ladder.

Primary -> Secondary -> [Lang fallback, Legal escalation, Tech-fault, Billing, Refund, Human-handoff]

80% of calls resolve at Primary; the ladder catches the long tail.

graph TD
    T[Triage Agent<br/>Sonnet 4.6] -->|buy| PS[Property Search<br/>Sonnet 4.6]
    T -->|sell| SI[Suburb Intelligence<br/>Sonnet 4.6]
    T -->|finance| MT[Mortgage<br/>Opus 4.7]
    T -->|compliance| CO[Compliance<br/>Opus 4.7]
    T -->|book| BK[Booking<br/>Sonnet 4.6]
    MT -->|escalate| HM[Human Mortgage Broker]
    CO -->|escalate| HC[Human Compliance Officer]

Migration / build steps

1. Map your intents. List the top 10-20 things callers ask for. Each becomes a candidate specialist.
2. Build Triage first. Triage is the smallest agent — only handoff tools, no domain tools. Get it right before building specialists.
3. Build specialists with focused tool surfaces. 5-8 tools per specialist is the sweet spot.
4. Wire context transfer. Each handoff carries a structured payload (intent, qualification state, prior tool outputs).
5. Instrument the delegation chain. Every conversation produces a handoff trace. Use it for debugging.
6. Add a human-escalation tool to every specialist. When uncertainty is high, escalate.

FAQ

How many specialists is too many? Above 10 handoff targets, the triage agent struggles. Group into a 2-level hierarchy.

Should each specialist be its own agent or one agent with a big system prompt? Separate agents. The mental-model and observability gains are worth the per-handoff overhead.

What is the latency cost of a handoff? ~200ms for context transfer plus the specialist's first model call. With streaming TTS the user does not perceive it.

Can specialists re-trigger triage mid-conversation? Yes. If a user pivots ("actually, I want to refinance, not buy"), the specialist hands back to triage which routes to Mortgage.

Where can I see this in practice? Our demo page has live examples for real estate and IT services verticals. Every 14-day trial tenant ships with this topology.

Sources

• The Handoff Pattern for Voice Agents - LiveKit
• Supervisor Pattern for Multi-Agent Voice AI - LiveKit
• OpenAI Agent Orchestration - Handoffs vs Agents as Tools