I built four agents to disagree with me

I’m the sole PM on a discovery project, which means there’s no product team to debate with, no design review, no engineering pushback. That sounds efficient until you realize those friction points are where bad ideas get caught.

So I built my own friction. I created four AI agents, each with a distinct perspective: one challenges my assumptions about what users actually said versus what I want them to have said, one pressure-tests the business case, one asks the uncomfortable questions I’m avoiding, and one models how a specific senior leader would react to what I’m proposing. They don’t tell me what to do. They ask the questions that role would ask. The decisions stay with me.

Before presenting an outcomes framework to leadership, I ran it through all four. The session surfaced concerns I hadn’t considered: success criteria that were too easy to hit, interview data that could be explained by politeness rather than genuine interest, a missing scenario for ambiguous results, and a framing that would trigger skepticism from a specific stakeholder. I addressed all of them before the real conversation. The meeting was smoother and I didn’t waste a senior leader’s time debugging problems I could have caught myself.

The most useful moment was when I was about to share a product demo with a stakeholder and all four agents, from completely different angles, said the same thing: don’t show it yet. The demo was built to close a decision, not to open a conversation. My instinct was to lead with the most impressive artifact, but showing it too early would anchor the discussion on execution details instead of the data that should drive the go/no-go call. I waited. It was the right call.

When I think through a problem alone, I tend to build the case for my preferred answer. The agents don’t have a preferred answer, they have a preferred set of questions. A solo PM who only debates with themselves will always win the argument. This makes sure someone in the room disagrees.