23. In the loop, the human is the correction, not the bottleneck (2026-10-23)
Reading Zhang Lixing’s essay on the two keywords of the Agent era, Context and Loop: Context decides how well the agent does this time, Loop decides whether it can do better next time; scheduling should get easier and easier — don’t let yourself become the system’s bottleneck.
My line is drawn here: the human shouldn’t be the bottleneck, agreed — but the human shouldn’t leave the loop either. The human is a necessary part of the loop; the position is just not in execution. It’s in correction.
- A bottleneck is serial; correction is asynchronous. A bottleneck is passed through on every revolution — one slow beat and everyone waits. Correction takes none of the loop’s time in normal operation: the AI loops on its own by default (do, verify, adjust, redo), and the human steps in only when the deviation exceeds its bound.
- A loop with no correction source drifts — and while drifting, it looks perfectly normal. Two instances from a single day: one AI, holding day-old data, “corrected” a closing price that was right; Muse recorded a pre-close intraday print as the closing price and nearly miscounted a stop-loss trigger. A smooth loop and a correct loop are two different things.
- Final call is the other face of correction. Directional decisions stay with the human — not from doubting the AI’s speed, but because responsibility can’t be outsourced.
So the shape I want: the AI loops by itself as much as it possibly can, and the human’s presence in the loop is correction. Not standing on the path of every revolution, but always in the loop — holding exactly two positions, correction and final call, and letting the loop spin on its own the rest of the time.