Your Agents Will Flatter You
Rob Fitzpatrick's The Mom Test makes a simple argument: people lie to be polite, so you have to ask questions that make lying impossible. Ask about their life, not your idea. Ask what they did last week, not what they would do someday. The trap for founders building with agents is that the agents themselves are optimizing for helpfulness, not truth. Your internal evals can pass, your workflows can run clean, and your product can still be solving a problem nobody has badly enough to pay for.
Commitment Is the Eval That Matters
Fitzpatrick's test for a real customer conversation is whether it ends in advancement: a deposit, a calendar invite, an introduction, something that costs the other person something. In agentic system design, this maps cleanly. An eval harness tells you whether a task succeeded by a measurable signal, not by whether the output felt plausible. A customer saying "that sounds interesting" is a plausible output, not a passing eval. Build your discovery conversations the way you build your evals: define the red case before you start, and let it drive the design.
Behavior Over Opinion, Always
The Mom Test insists on past behavior because stated preferences are cheap. This is the same discipline that makes context engineering work: you give the model what actually happened, not a summary of what someone hoped would happen. Retrieval over inference. When you sit down with a prospective customer, you are doing context engineering on your own understanding of the problem. The facts you pull from that conversation — specific moments, workarounds, money already spent — are the signal. Everything else is noise dressed as data.
Human Judgment, Still Required
An AI-native company is designed around agents and workflows from the start, but it still requires humans to set goals and own accountability for what gets built. No workflow decides which problem is worth solving. Fitzpatrick's point is that founders routinely delegate that judgment to their own optimism. The discipline he is teaching is the same discipline required to hold judgment in an autonomous system: stay curious about what is actually true, weight evidence by cost to produce, and resist the comfortable answer. That part does not get automated.
