What 'multi-agent pipeline' actually means in the average enterprise demo

Every few weeks someone sends me a demo video. Confident exec on stage, an agent that handles intake, routes the request, calls the right tool, drafts a response, and closes the loop. End-to-end automation. The future of work, basically. I always want to ask the same question: what's running on a Tuesday afternoon when the demo team isn't watching?

Here's what the architecture usually looks like once you get past the slide:

• Retrieval: A vector store with your company docs, chunked at whatever size didn't break the coherence too badly. Works great on the questions it was tested with. Struggles the first time someone asks something that lives across three documents that were never meant to be read together.
• Tool calls: One or two API calls dressed up as 'the agent taking action.' Booking a calendar slot. Pulling a row from a CRM. Real and useful. Not the same as autonomous decision-making.
• Prompt engineering: The actual product. Three months of iteration, a system prompt that's now 800 tokens long, a dozen edge cases handled by increasingly baroque instructions. The model is smart. The prompt is holding the shape.
• The human in the loop: Not a failure of ambition. The right call. For anything consequential — a customer-facing response, a contract clause, a clinical flag — someone is reviewing before it ships. The demos don't linger on this part.

None of this means the system doesn't work. Some of these deployments are genuinely moving metrics. The demo video didn't lie. It just skipped the part where the intern checks the output before it goes to the customer.