AI Agents
Coding Agents Are Fast. Taste Is the Bottleneck.
Generation got cheap. Deciding what should exist, and when to stop, did not get any cheaper.
The interesting constraint in building software with an agent is no longer writing the code. Generation is effectively free. What is scarce is deciding what should exist.
I have been building against this for a few months now — a coding assistant in the terminal, plus a lot of smaller agent-shaped tools. The pattern is consistent enough to write down.
Speed moves the bottleneck upstream
When implementing took a week, the design question got answered slowly and mostly implicitly, by whatever the first reasonable implementation happened to be. When implementing takes an hour, that question gets answered fast — which means it has to be answered deliberately, because the default is to let the first plausible thing through.
The bottleneck moved from typing to judgement. Most workflows did not notice.
An agent will happily implement your first idea. It will not tell you the first idea is wrong, because it has no way to know what you were actually trying to do.
Taste is a stopping rule
What experienced engineers have, concretely, is not a better sense of what is possible. It is a better sense of when to stop. That is a stopping rule, and it is learnable, but only if you are willing to be wrong sometimes in order to calibrate it.
Concretely: knowing that the third abstraction is premature, that a second cache is not an optimisation, that a 600-line component wants to be two files. None of those are things an agent can infer, because in each case the code works.
What this changes about review
Reviewing generated code line by line does not scale and was never the point. The valuable review question shifts upward: should this exist, is this the right shape, and is there a simpler version that does the same thing.
I find the practical version of this is to write the shape down before asking for the implementation. Two sentences about what the module owns and what it does not. It removes most of the wrong-direction work, because the agent has something to disagree with.