Orange Business runs AI agents for 120,000 staff from an agent registry and an MCP registry. Small teams need the same list

When automations multiply, the problem stops being building them and becomes knowing which ones exist and what they can touch.

What happened

LangChain published a case study on Live Intelligence, Orange Business's AI platform. It is deployed across 120,000 Orange employees and adopted by more than 150 customers in France and Europe. Orange evaluated roughly 20 vendors and chose LangChain for model neutrality and portability, with LangSmith for evaluation and observability. Its head of AI products, Mathieu Ducrot, said building that observability in house would likely have delayed the roadmap twelve to eighteen months.

Two registries sit at the center. An Agent Registry lets teams discover, reuse, version and govern agents and prompts. An MCP Registry is a central catalog of trusted tools and connectors, so agents reach Microsoft 365, CRMs and internal APIs through a controlled, auditable layer. Prompt versions are kept per environment so changes are tested before production.

Orange also runs three evaluation layers: RAG quality with golden datasets, agent behavior including tool selection and prompt injection robustness, and production monitoring of latency, cost, failures and escalations. Regression evals run before changes ship.

My take

You are not Orange, but the failure it is avoiding shows up at ten people too. Most small businesses I look at have automations nobody fully remembers: a Zap someone's former VA built, an n8n flow with an owner's personal API key, a GoHighLevel workflow that still texts an old offer.

A registry can start as one sheet. For each automation or agent: what it does, who owns it, which accounts and keys it uses, what it can write or delete, and when it was last tested. Do the same for connections: which CRM, inbox and payment tools are approved, and with what scope.

Then copy the cheap part of their eval setup. Keep ten real examples per agent, run them before any prompt change, and review failures and escalations monthly. That list is what lets you change a model, rotate a key or hand over a client without breaking something quietly.