If your agents are exposing real friction — tasks they can't quite close, handoffs that keep breaking — that's not a bug to suppress, it's a product idea to follow. Paul Graham's logic applies harder when the tooling is this new.
Paul Graham's essay on finding startup ideas is, at its core, an argument against deliberate invention. The productive move isn't to brainstorm what the market wants — it's to live close enough to a real problem that you can't ignore it anymore. Ideas that emerge this way have a specific texture: they seem narrow, they seem almost too simple, and the people who want them want them badly enough that a rough version will do. That desperation in early users is the signal worth chasing.
The translation is almost uncomfortably direct. If you're running agents in production, you have an unusually honest view of where workflows actually break — not where people say they break in a customer interview, but where the automation quietly gives up and routes back to a human. Those handoff points are Graham's live problems, surfaced in real time. The founders who will find the sharpest ideas right now are the ones treating their own agent infrastructure as a source of embarrassing friction worth solving, not papering over.
- Build for the pain you personally can't route around
- trust early users who want the rough version right now
- the place your agents fail most often is probably a company.
