Merck and Mastercard Built the Plumbing First. Your Agent Needs It Too.
Merck cut discovery cycles by a third and marketing reviews by 80%, but only after building the backend first. For founders shipping AI agents, the lesson is simple: the agent is only as good as the pipes underneath it
The Agent Is the Sizzle. The Plumbing Is the Steak.
Merck just cut one drug discovery cycle by a third. Mastercard is shaving marketing review times by seventy to eighty percent. The headlines will credit AI agents, but the people inside both companies are telling a different story. Sean Finnerty, Merck's VP of Digital Platforms, says the only reason any of this works is because they built the infrastructure first. The models didn't magically get better at pharma law. The backend started enforcing rules, routing drafts through review queues, and keeping audit trails intact. The intelligence lives in the workflow, not the LLM alone.
Why Plumbing Breaks First
Most AI demos look fantastic in a sandbox. Then you wire them to a real database and everything slows down or falls apart. Context windows fill up. APIs time out. Vector searches return irrelevant chunks. Long-running tasks die halfway through because the serverless function hit its timeout. These aren't model failures. They are backend failures wearing model makeup.
Your agent needs to read state in real time, write without race conditions, and remember what it was doing after a user closes their laptop. If the database can't stream updates or the workflow engine can't recover from a crash, you don't have an agent. You have a chatbot with delusions of grandeur.
Big Company Lessons for Small Teams
Merck and Mastercard can afford to hire fifty engineers to build custom orchestration layers and bespoke memory systems. Indie hackers, solo founders, and small product teams cannot. They need the plumbing already assembled so they can spend their limited time on user experience and business logic instead of patching message queues at two in the morning.
This is exactly why the stack matters. A reactive database that pushes updates to the frontend changes how an agent feels to use. Durable workflows that survive restarts keep long tasks from evaporating. Native vector search turns memory from a fragile integration into a simple query. Botflow runs on Convex and wires these pieces together so you skip the six-month infrastructure phase and start shipping.
You do not need a six-figure infrastructure budget to get enterprise-grade reliability. You need the right primitives in place from the start.
Build the Pipes, Then the Personality
Founders love to obsess over model choice and prompt engineering. Those are the visible knobs, so they get the attention. But they are the last ten percent. The first ninety percent is making sure your agent can read, write, remember, and recover without a human babysitting every step.
If you are shipping an AI product this week, be honest about one thing. Your backend either supports the agent, or the agent is carrying your backend on its back. When it is the latter, your users will feel it in the lag, the dropped tasks, and the weird amnesia that sets in halfway through a session. Fix the pipes first. The flashy stuff can wait.