Fowler and Jain measure what scaffolding costs to run an AI coder

Martin Fowler and Rouan Jain ran a controlled experiment on the automated checks that verify an AI coding agent's output. Their answer: yes, build them, but cost and reliability decide whether the system holds up. Checks with weak design can cost more to run than the agent work they validate. Unreliable checks produce false signals that erode trust in everything the agent produces. The experiment produced concrete numbers. Running a full set of checks on each agent step was expensive. Running a lighter check more often and a heavier check less often cut costs without sacrificing confidence. Fowler and Jain found that check design, meaning what to verify, how often, and at what depth, is a first-class engineering decision. A team that treats check design as secondary ends up overspending or working without reliable feedback. The checks you build around an agent shape how much you can trust its output and what that trust costs. Designing those checks is design work. A reviewer depending on noisy or expensive checks loses a clear signal before they ever see the agent's result.