If agents do the deep work now, the interruption tax hasn't disappeared — it's just moved somewhere founders rarely look. Getting this wrong quietly caps what your stack can actually produce.
Paul Graham's essay draws a clean line between two incompatible operating systems for time. Managers slice the day into one-hour appointments and lose little when a meeting lands anywhere on the clock. Makers — programmers, writers, anyone building something coherent — need half-day minimums because a single meeting doesn't cost one hour, it costs the entire surrounding block. The damage is asymmetric and mostly invisible, which is why the people running makers rarely feel the full weight of what they're doing to output.
The maker in your stack is now the agent, but the scheduling problem is yours. Every time a human approval gate, a Slack interrupt, or a poorly timed review cycle lands in the middle of a long agent run, you've replicated the exact pathology Graham described — except the maker can't tell you it's annoyed. Founders building agent-native workflows need to treat uninterrupted execution windows as infrastructure decisions, not calendar courtesies. Batching human judgment into designated checkpoints rather than sprinkling it across the day is how you stop being the meeting that wrecks your own maker's morning.
- Design agent workflows around execution windows, not human availability windows
- batch human-in-the-loop moments the way Graham batches meetings — to one end of the day
- the cost of an interruption to a running agent process is a design flaw, not a scheduling inconvenience.
