The Challenge
QVC Group wanted to move beyond scripted chatbots to deliver genuinely intelligent customer experiences - agents that could reason about context, access tools, and take meaningful actions. This was in the early days of LLM-based agents, before frameworks like MCP existed and before "agentic AI" was an industry buzzword. We had to build from raw materials, digging into the core DNA of what makes an agent truly autonomous.
My Approach
I designed and shipped production AI agents across two markets, each embedded within major QVC content initiatives.
QLINARISCH - QVC Germany
QLINARISCH is QVC Germany's culinary content platform - a 360-degree experience featuring recipes, livestreams, and chef partnerships. Inside it, QUINNY serves as an AI agent that goes beyond simple Q&A: it can search recipe databases, reason about ingredient substitutions, navigate product catalogs, and provide contextual recommendations - all orchestrated through tool-use patterns built on LangChain.
My Garden Escape - QVC UK
My Garden Escape is QVC UK's gardening content hub, featuring daily live shows, expert advice, and seasonal planning. The ASK ME agent powers intelligent interactions - answering gardening questions by accessing plant databases, seasonal calendars, and product inventories through structured tool calls, not pattern-matched scripts.
Architecture
Both agents share a common architecture pattern:
- LLM backbone: Large language model for reasoning and natural language understanding
- Tool orchestration: Custom tool-use framework built on LangChain, giving agents structured access to databases, APIs, and search capabilities
- Context management: Conversation state and user context tracking for multi-turn interactions
- Guardrails: Output validation, content safety filters, and fallback paths for edge cases
This was built before the Model Context Protocol (MCP) standardized tool access - we had to design our own tool interfaces, serialization formats, and orchestration patterns from scratch.
Key Decisions & Trade-offs
Custom tool interfaces over generic plugins: Without MCP, we built bespoke tool connectors for each data source. More work upfront, but it gave us precise control over latency, error handling, and response quality.
Agents, not chatbots: The key distinction is tool use - these agents don't just generate text from a prompt. They reason about which tools to invoke, interpret structured results, and compose multi-step answers. This was a deliberate architectural choice that required significantly more engineering than a retrieval chatbot.
Market-specific tuning: Each agent is tuned for its market - QUINNY operates in German with culinary domain knowledge, ASK ME in English with horticultural expertise. Same architecture, different specializations.
Impact
Both agents have been running in production for over three years, serving real customers at scale across two European markets. They demonstrated the viability of production agentic AI in enterprise retail long before the industry converged on agent frameworks - and the architecture patterns we pioneered directly informed the agentic systems we're building today.