Software is heading toward the dark factory: agents doing the implementation, humans holding
the two endpoints.
My job: write specifications an agent cannot misread, and demand receipts it cannot fake.
When agents can produce working code from any sufficiently precise description, the bottleneck moves upstream to the description. The specification becomes the primary engineering artifact; the codebase is a derivative — closer to a build output than to source.
Agent failures are ambiguity failures of the loop, not intelligence failures of the agent. Smartness cannot supply a fact that was never specified. A real run has a goal, a boundary, tools, artifacts, and receipts — miss one and you made a wish, not a delegation.
An agent declaring success is self-attestation by the party with the strongest incentive to call the job done. No diff, no test run, no artifact — no "done." The same review bar applies whether the author was a human or a machine.
Non-deterministic output demands a deterministic, accessible interface as its stability layer. The UI is a contract, not a display — that is what makes streaming AI trustworthy enough for 50M+ monthly users on a financial news platform.
AI equalizes execution speed — but execution was already cheap. What it amplifies is specification quality, which is a direct function of domain depth. It makes experts more productive faster than it makes novices competent. Keep earning the depth.
"I use AI" stopped being the differentiator. Whether you can hand an agent a boundary and a review bar — instead of still typing every line yourself — is the one that's left.
— from my claim vault, on where the leverage moved