Ask an AI tool the same question two weeks in a row and you might get two completely different answers. Anant Gupta (CTO @ Mixpanel) says that is not a model problem. It is an architecture problem.
In this episode of 1 IDEA, Suresh Mathew sits down with Anant to unpack what happened when an agent told him Mixpanel had three times the users his own dashboard showed, and what has to sit between your data and your agent before you can trust the answer.
We cover:
- Why the right context outperforms a better model
- How to make your most critical answers reproducible with code
- Why AI agents should write the code instead of making the tool calls
- Why AI usage is up while business impact stays flat
CHAPTERS
00:00 Introduction
01:04 The Agent That Gave The Wrong Number
03:10 A Context Problem, Not A Model Problem
04:50 Building The Context Engine
09:12 Why Reliability Is A Step Function
10:44 Giving Customers The Choice On AI
22:33 Mixpanel Headless: Solving Determinism With Code
31:30 Why AI Usage Is Up But Impact Is Not
39:56 Verified Mode And Trusting Your Data