If you're deploying agents on complex tasks, how you schedule human checkpoints is directly shaping what those agents can actually finish — and most founders haven't thought about it once.
Paul Graham's essay draws a hard line between two operating modes. Managers live in one-hour increments, hopping meeting to meeting with no friction cost. Makers — writers, programmers, anyone doing generative work — need half-day blocks at minimum, because a single meeting dropped into the middle of an afternoon doesn't just consume an hour, it consumes the whole afternoon. The damage is architectural, not arithmetic. Graham's point is that schedule design is output design, and most organizations default to the manager's schedule because managers run the organizations.
Agents are the new makers on your team, and they inherit the same vulnerability. A long-horizon research task, a multi-step code generation run, an autonomous drafting pipeline — these need uninterrupted context windows and clean handoff conditions. When you drop a human-review checkpoint into the middle of an agent run because it fits your calendar, you're imposing the manager's schedule on a maker's process. The design question Graham forces on human teams now applies to your orchestration layer: where are the seams, who called them, and did output suffer because of it.
- place human-in-the-loop checkpoints at natural task boundaries, not calendar convenience
- audit which interruptions serve trust and which just serve habit
- treat your agent's working context as the scarce resource Graham always said the maker's afternoon was.
