Building
Agents Do Not Remove the Work, They Move It
The typing got faster. Deciding what to build, and being able to tell when it is wrong, got more important.
The current narrative around AI coding agents is that software development has been automated away: you prompt the model, sit back, and a complete, production-ready product emerges thirty seconds later.
Anyone who actually ships commercial software knows this is nonsense. Agents did not remove the work; they moved the work upstream.
The shift from typing to architecture and verification
When writing code manually took forty hours, the design decisions happened slowly and incrementally along the way. You had time to notice architectural flaws as you typed each function. When an agent generates three hundred lines of code in four seconds, that mental cushion disappears.
- Specification clarity: if your instructions are ambiguous, the agent will confidently generate the wrong implementation at lightning speed.
- System boundary enforcement: agents love to introduce dependencies and expand scope silently unless strictly constrained.
- Verification rigor: you must have automated tests, strict types, and verification harnesses that can reject flawed pull requests.
Agents accelerate execution. They do not supply taste, judgement, or stopping rules.
The scarce skill is no longer knowing the syntax for a React hook or SQL query. The scarce skill is knowing which abstractions should exist in the first place, and having the discipline to delete the plausible-looking code that does not belong.