Scaling Agents to Millions of Calls/Week: Customization Without Chaos

​Every new enterprise customer wants something different, and every customization makes your product harder to maintain. So how do you build one agent platform for dozens of clients, each with its own rules, without ending up with dozens of versions of your code?

​Lets bring it down to earth: Take a cancelled appointment. Happy Clinic wants the agent to offer the slot to the next caller right away. Sunny Hospital wants the same slot held for the doctor's waitlist. Both are paying customers, and both are right about their own rules. Do you add a flag, or say no?

​In our first tech talk, we showed how we turn probabilistic models into deterministic outcomes. This time we look at what happens when those rules multiply. Today we run dozens of clients with hundreds of customizations across more than ten data systems, at millions of calls a week, from one codebase.

​The easy answers break at scale. Add a flag and every function turns into a wall of "if client == ...". Say no and you lose the deal. We'll show what we did instead. AI PMs and Forward Deployed Engineers ship client-specific code straight into the core product, and the architecture keeps that from turning into chaos.

​Agenda

​- Prosper Pitch + Business/Product Context: 5 minutes
- Tech Talk: 20 minutes
- Why agents are hard to build, and why that needs AI PMs
- The role of the FDE
- Custom code vs platform: who decides, and when
- How we tame the beast of custom code
- Open Q&A: 10 minutes
- Networking & Beers