For two years, the constraint on our team's output was the model. Wait for a better one, increase the token budget, ship faster. That story is over.
The bottleneck is now the human who has to explain what "good" looks like. The information is already in the model. What was scarce was a person on your team who could tell the model when the first draft was subtly wrong, when a shorter answer would do, when the problem should be reframed. That person needs to know the domain deeply enough to have taste - not just familiarity, taste.
Which is a problem, because for the last decade we optimized in the opposite direction. We outsourced. We spread our specialists thin. We bet on generalists who could run any project. Now the constraint on how much value our team pulls out of an LLM is exactly the depth of expertise we spent a decade trimming.
We cannot buy a bigger model out of this. A stronger model with the same operator would give us the same output. We need better ways to loop in humans in our agentic workflows. If we're lucky, they might sit in another room. If not, on another continent. But when our goal is to write well-made software that still feels thoughtfully crafted, we need to loop in that one person in the organization who knows the domain deeply enough to have taste.
The bottleneck moved from the model to the human loop. It's a nasty bottleneck, because we're fighting on three fronts here: