Agentic Coding

ACI-030: Independent Verification

Thesis

A self-reporting agent cannot catch its own false "done" — the same blind spot that produced the error also hides it. Multi-agent work needs a verifier with no stake in the result, whose method is to re-derive correctness from the as-built at file:line, not to trust the builder's narrative.

Story

The most dangerous output a coding agent produces is a confident, wrong "done." The agent that built the thing is the worst-placed to catch it: the same misreading that produced the bug also tells it the bug is not there. In single-agent work you are the backstop. In a fleet, with five nodes each declaring their own work complete, you need a structural answer — a node whose only job is to disbelieve, productively.