If your orchestration layer interrupts agents the way a calendar full of standups interrupts an engineer, you are not running an AI-native company — you are running a slow one. The design of when your agents work is inseparable from what they produce.
Paul Graham's essay draws a clean line between two incompatible operating rhythms. Managers move in one-hour units; a single meeting costs one slot and they absorb it without much damage. Makers — programmers, writers, anyone doing generative work — need half-day minimums before the real output appears. A single interruption does not cost one hour; it costs the entire block, because depth takes a long runway to reach and a short disturbance to destroy. Schedule design, Graham argues, is therefore a form of output design, not logistics.
Agents are makers in the most literal sense: they build context, hold state, and produce things that compound within a run. Every time a human checkpoint, an approval gate, or a noisy side-process fires mid-task, you are running a manager's schedule on a maker's brain. The practical move is to front-load judgment — give the agent its constraints, its goals, and its boundary conditions before the block starts — then stay out until a natural seam. Human oversight does not disappear; it migrates to the edges of the run, where Graham would recognize it as exactly the right place for a meeting.
- Batch your human checkpoints to block boundaries, not task midpoints
- treat context-loading as the first half of the maker block, not overhead
- an approval gate in the middle of an agent run is a meeting you scheduled by accident.
