Discovery Before Deployment
Cagan's empowered teams run discovery in parallel with delivery — they are always asking whether the next thing is worth building, not just whether it can be built. The agent equivalent is building your evaluation harness before you build the agent. If you cannot specify what good output looks like in measurable terms, you have not yet done your discovery. Red cases are not a testing afterthought; they are the product definition.
Outcome Over Output, Applied to Workflows
Cagan's central complaint about feature factories is that they measure shipping, not results. The same trap appears in agentic systems when founders count tasks completed instead of outcomes reached. Before graduating a workflow to a true agent — one that chooses its own path — you need a clear outcome metric and evidence that the branching complexity is real. Most automation problems are workflow problems dressed up as agency, and a workflow with a good eval is strictly better than a loose agent with none.
Where Human Judgment Holds the Line
Cagan insists product teams own accountability for results, not just execution. Autonomous systems do not dissolve that ownership. A founder still must set the goals an agent pursues, define the boundaries it cannot cross, and stay accountable when it fails. The eval harness is how you make that accountability operational rather than rhetorical — you cannot own outcomes you cannot measure.
