Agents Inherit Your Assumptions
Every workflow you build encodes a belief about how customers behave. If that belief came from a conversation where someone said "oh, that sounds great," your agent is running on flattery. Fitzpatrick's core move is to replace opinion-fishing with behavior archaeology: what did they do last time, how often, what did they try first, what did they pay. Concrete past behavior is the only input an agent can actually act on without compounding the original distortion.
Compliments Are Noise; Commitment Is Signal
A customer who calls your idea clever has told you nothing an eval can use. A customer who forwards you their current spreadsheet, agrees to a pilot, or pays a deposit has given you structured data. Fitzpatrick frames commitment and advancement as the real test of a conversation's worth. For a founder building agentic tools, this translates directly: design your discovery process so the output is a measurable action, not a sentiment. That output becomes context the agent can actually work with.
Design the Interview Like You Design the Eval
Fitzpatrick argues you should know what a good conversation looks like before you have it. That is identical to the discipline of writing evaluation criteria before shipping an agent. If you cannot state in advance what a successful customer conversation would surface, you cannot tell afterward whether it succeeded. Build the rubric first — for the interview and for the agent — and let the hard, disconfirming cases drive both designs.
Human Judgment Owns the Interpretation Layer
Agents can surface patterns across dozens of customer interviews, flag recurring complaints, and cluster signals. What they cannot do is decide whether a pattern is real or an artifact of who you talked to, or whether a commitment signal reflects genuine need or politeness. That call belongs to a human. The Mom Test discipline and agentic tooling are most powerful together when the founder holds judgment over what the data means and the agent handles the accumulation and retrieval of what was actually said.
