Inside the CallSphere Hotel Stack: 11 Specialist Agents + Tools

TL;DR

CallSphere's hotel platform is built on OpenAI Agents SDK + Realtime API with 11 specialist agents, each with dedicated tools and handoff logic. Here is the full technical architecture.

Framework Choice

• Voice: OpenAI Realtime API (GPT-4o-realtime-preview)
  • PCM16 24kHz audio
  • Server VAD turn detection
  • Sub-1-second latency
• Orchestration: OpenAI Agents SDK (hierarchical handoffs)
• Analytics: GPT-4o-mini for post-call sentiment, intent, and summarization
• Telephony: Twilio (voice) + SIP trunking
• Backend: Python FastAPI + NATS message queue
• Storage: PostgreSQL (primary), Redis (cache), ChromaDB (RAG)

The 11 Agents

1. Concierge Agent (Triage)

• Model: GPT-4o-realtime
• Tools: lookup_guest_by_phone, detect_intent, get_hotel_info, route_to_agent
• Handoffs: all other agents

2. Reservation Agent

• Tools: search_availability, quote_rate, check_parity, collect_guest_details, process_deposit, create_reservation
• Integrations: PMS, channel manager, payment processor

3. Check-In Agent

• Tools: verify_reservation, capture_incidentals, issue_mobile_key, assign_room
• Integrations: PMS, Salto/Assa Abloy/Dormakaba, Stripe

4. Check-Out Agent

• Tools: pull_folio, dispute_charge, capture_payment, email_receipt, post_loyalty_points

5. Housekeeping Agent

• Tools: update_room_status, create_maintenance_ticket, assign_cleaner, report_inspection

6. Guest Services Agent

• Tools: create_service_request, schedule_wake_up, book_restaurant, extend_check_out, book_activity

7. Group Sales Agent

• Tools: check_block_inventory, draft_proposal, schedule_site_visit, notify_dosm

8. Revenue Signals Agent

• Mode: Passive observer, not guest-facing
• Tools: analyze_demand, check_parity, alert_revenue_manager

9. OTA Channel Agent

• Tools: push_rates, sync_inventory, handle_overbooking
• Integrations: Siteminder, Cloudbeds, Derbysoft

10. Loyalty Agent

• Tools: lookup_loyalty, apply_discount, trigger_upgrade, log_preference
• Integrations: Marriott Bonvoy, Hilton Honors, etc.

11. Night Audit + Emergency Agent

• Tools: late_check_in, verify_identity, classify_emergency, escalate_to_on_call
• Runs: 12 AM–7 AM autonomously

Handoff Logic

Handoffs use OpenAI Agents SDK's hierarchical handoff mechanism. Each handoff carries:

flowchart LR
    CALLER(["Guest or Prospect"])
    subgraph TEL["Telephony"]
        SIP["Twilio SIP and PSTN"]
    end
    subgraph BRAIN["Hotel Concierge AI Agent"]
        STT["Streaming STT<br/>Deepgram or Whisper"]
        NLU{"Intent and<br/>Entity Extraction"}
        TOOLS["Tool Calls"]
        TTS["Streaming TTS<br/>ElevenLabs or Rime"]
    end
    subgraph DATA["Live Data Plane"]
        CRM[("CRM and Notes")]
        CAL[("Calendar and<br/>Schedule")]
        KB[("Knowledge Base<br/>and Policies")]
    end
    subgraph OUT["Outcomes"]
        O1(["Reservation confirmed"])
        O2(["Room service order"])
        O3(["Front desk handoff"])
    end
    CALLER --> SIP --> STT --> NLU
    NLU -->|Lookup| TOOLS
    TOOLS <--> CRM
    TOOLS <--> CAL
    TOOLS <--> KB
    NLU --> TTS --> SIP --> CALLER
    NLU -->|Resolved| O1
    NLU -->|Schedule| O2
    NLU -->|Escalate| O3
    style CALLER fill:#f1f5f9,stroke:#64748b,color:#0f172a
    style NLU fill:#4f46e5,stroke:#4338ca,color:#fff
    style O1 fill:#059669,stroke:#047857,color:#fff
    style O2 fill:#0ea5e9,stroke:#0369a1,color:#fff
    style O3 fill:#f59e0b,stroke:#d97706,color:#1f2937

• Full conversation history
• Guest profile context
• Current PMS state
• Previous agent notes

RAG for Policy Questions

Each property ingests policy documents (cancellation, pet, parking, amenity, etc.) into ChromaDB. Agents query ChromaDB for accurate policy citation instead of hallucinating.

Guardrails

Multiple guardrails prevent hallucination:

• Input validation (language detection, intent confidence)
• Tool call validation (schema enforcement)
• Output validation (policy accuracy checks)
• Hallucination detection (comparing claims to RAG ground truth)

FAQ

Q: Is this open source? A: No. CallSphere is a managed SaaS.

Q: Can I use my own LLM? A: On enterprise plans, Claude and Gemini available as alternatives.

Q: What's the latency? A: <1 second first response, <200ms for tool calls.

---

Related: 11-agent stack overview | Hotel industry

#Architecture #MultiAgent #Technical #CallSphere