Markus Eisele Says AI Coding Agents Need 3-Step Specs and Validation

1 articles · Updated · O'Reilly Media · Jul 8

Eisele argues AI coding agents do not eliminate specification work; they shift costs into correction loops, human review and rework when requirements are vague.
A U-shaped cost curve makes zero-spec prompting look cheap only upfront, he says, with the lowest total cost usually near structured acceptance criteria or BDD scenarios rather than 0 specification.
He adds a missing middle step: validate the spec before implementation, checking for contradictions, gaps, untestable claims and cases where the written requirement misses the real intent.
A 3-step workflow—agent drafts the spec, another agent attacks it, then implementation agents code against human-approved tests—can cut effort without removing the need for review.
For multi-agent pipelines, he says the sweet spot shifts toward tighter contracts, schemas and executable validators because drift compounds across handoffs and turns one flawed spec into parallel errors.