The Agentic CX Stack: Architecture Overview

AI agents are reshaping how customers interact with brands. Instead of clicking through menus and filling out forms, customers simply say what they need, and an intelligent agent handles the rest — searching products, checking accounts, processing requests, completing transactions.

But enabling this experience requires more than wrapping existing APIs with an LLM. This post walks through the complete reference architecture we use when designing agentic CX systems for enterprise environments.

The Architectural Layers

The agentic CX stack separates into five clean layers, each with clear responsibilities and well-defined interfaces:

Experience Layer: Multimodal AI

Two capabilities power the customer-facing experience:

Gemini Live API enables real-time multimodal conversations — high-quality streaming audio, natural speech synthesis, real-time video frames for visual context, and bidirectional text messaging. Customers can show a product via webcam, speak naturally, and receive semantically relevant responses — all within a single conversation.

Vertex AI Search provides semantic product and knowledge discovery — understanding natural language queries, visual similarity, and customer intent, not just keyword matching.

Agentic Layer: The Digital Concierge

Built on the Google Agent Development Kit (ADK), the agentic layer is where intelligence meets action. An ADK Orchestrator classifies user intent and routes to specialized agents:

• Service intents → Service Agent (account management, status, resolution)
• Transactional intents → Transaction Agent (orders, returns, scheduling)
• Escalation intents → Human handoff with full context

When switching between agents — or handing off to a human — all relevant context travels with the conversation. The customer never has to repeat themselves.

Tool Layer: MCP for Unified Execution

The Model Context Protocol (MCP) provides unified tool discovery and routing. Tools are registered with a central registry and discovered by agents at runtime. When you add a new capability, you don't modify the agent — you register a new MCP tool, and any agent can discover and use it immediately.

This dramatically reduces integration cycles. Platform teams evolve capabilities without retraining or redeploying agents. MCP servers typically include: search and knowledge retrieval, CRM and account access, transaction and order management, inventory and scheduling, and backend system connectors.

Open Standards: No Vendor Lock-In

The architecture is built on open interaction standards that form its governance fabric:

MCP (Model Context Protocol) — standardized tool access. Agents query a registry of available tools, understand their capabilities through semantic descriptions, and invoke them through a consistent interface.

A2A (Agent-to-Agent Protocol) — enables agent discovery and delegation. Each agent publishes an Agent Card describing its skills and capabilities, enabling dynamic routing between specialized agents.

These standards mean no vendor lock-in. You can swap implementations, add new integrations, and evolve your stack while maintaining interoperability across the ecosystem.

Collaboration Layer: Human Expertise in the Loop

AI handles most interactions. When judgment is needed — policy exceptions, complex empathy situations, high-value decisions — the architecture makes it easy to bring humans in without disrupting the customer experience.

Key patterns include:

• Human Escalation — full handoff to a human agent with complete conversation context
• Consultation (HITL) — AI consults a human expert mid-conversation, then continues. Used for policy approvals, exception handling, and high-value decisions.
• color: var(--text-primary);">Contact Center Integration — Voice bridges (WebRTC to SIP), context passing, and queue-based routing to specialized human agents

The architecture supports direct integration with Genesys Cloud, Twilio Voice, Cisco Webex Contact Center, and other CCaaS platforms via standard SIP/RTP protocols.

Key Design Principles

What makes this architecture distinctive — and what we apply consistently in enterprise engagements — is its treatment of human–AI collaboration as a deliberate design principle, not a fallback mechanism.

AI agents handle routine and high-volume interactions at scale. Human agents are engaged intentionally — for judgment, empathy, policy interpretation, or exception handling. The result is not just increased automation, but a more balanced operating model that delivers better customer experiences, improves operational efficiency, and empowers human agents to focus on meaningful, high-value work.