AI moved the work into the 10%
AI can make the first 90% feel magical. The remaining 10% turns the builder into a reviewer, tester, and manager.
AI changes where the effort lives.
When it works, it compresses days of implementation into minutes. When it misses, the work becomes an evidence hunt: find the invisible assumption, reproduce the failure, tighten the instruction, and check what else moved.
That is why managing AI agents can feel less like automation and more like managing a very fast remote worker. The loop is still review, feedback, review, feedback. The difference is speed—and the fact that the failure may be plausible enough to escape notice.
The opportunity is not pretending the broken 10% will disappear. It is building better feedback loops around it: explicit outcomes, small scopes, observable behavior, tests, version control, and a human who understands what “good” looks like.
AI is an extraordinary lever. Judgment is still the fulcrum.
Research trail: My note on managing AI agents and a longer conversation about AI-assisted web development.