The model isn't the moat, the harness is
Curly quotes broke the build. That's the part that stuck with me.
OpenAI's frontier team shipped a million lines with no human writing code, and the thing keeping quality up wasn't smarter agents. It was tests. Every taste call the design team made — typography, component rules, doc linking — got compiled into CI. Agent violates it, build goes red, agent fixes itself. Nobody had to review a diff to catch it.
The reframe I keep coming back to: the model isn't the moat, the harness is. You don't control how good GPT gets. You do control how precisely you've encoded your bar. That's the real skill now — managing a cluster of eager interns who need very clear rules.
Which changes what's scarce. Not building, but knowing what "good" means well enough to make it testable. That gap is where I've been pointing SoloVault, which digs up buildable market signals so the decision comes first.
How do you encode taste your agents can actually check?
Replies