Note / 05
Why agent verifiers fail
A second model does not automatically create independent assurance.
“Have another agent verify it” sounds like a clean solution. Sometimes it helps. It is not independence by default.
The implementer and verifier may share the same blind spots, context, tool failures, or incentive to produce a tidy answer. A verifier can confidently approve evidence it cannot reproduce. Two agents can agree because both inherited the same mistaken premise.
Stronger verification is heterogeneous:
- deterministic checks for crisp invariants,
- independent queries for important evidence,
- model judgment for semantic quality,
- calibrated evaluation against expert decisions,
- and human review when consequence or uncertainty crosses a threshold.
The objective is not more votes. It is failure modes that do not all collapse together.