· Long-running agents crash because the plan lives inside the model's head, which forgets. Move it into deterministic code, the outside plan, and treat the agent like dice on a boring table: replayable, scalable, and crash-proof.
· You don't flow with a compiler because you watch it well. You flow because you stopped watching it. The craft of running agents is making their work boring enough to ignore.
· We are automating the making and keeping the judging, on the assumption that judgment survives without the making. It doesn't, and the eye that only reviews starts grading the wrong thing.
· Agents paint fast and paint everything. The Renaissance workshop knew which strokes a master never gives away, and we are handing them off first.
· Why AI flattens org charts, and the judgment shortage that flattening quietly creates.