Home / Several agents and adversarial checking

Adversarial verification, judge panels and completeness critics

Three verification roles serve different angles and do not substitute for one another: the verifier subagent receives only the artefact to judge, the success criteria and the tools to check it, never the diagnosis or the history of whoever built it, and its silence is only legitimate if it genuinely searched and found nothing, since the default bias facing doubt leans towards rejection; a panel comparing several scored solutions is a widespread industry practice, not a pattern Anthropic names; and a completeness critic, a pattern documented under that exact name, looks exclusively for what is missing against the brief, never for errors in what was delivered.

A verifier subagent, a pattern documented by Anthropic, is an agent whose sole responsibility is to test or validate another agent's work. It receives the artefact to judge, clear success criteria and the tools needed to check it, without needing to understand why that artefact was built that way, nor to know the diagnosis its author already suspected before launching the verification.

An empty report is not a default report

A verifier that finds nothing can legitimately say so, but only if it genuinely searched and counts what it checked. The typical instruction the documentation proposes for this role leans the other way, towards rejection, in case of doubt: facing uncertainty, the agent should conclude with a refusal rather than with silence. An empty report is therefore not a default answer to uncertainty, it is only acceptable as the conclusion of an active, counted search.

Artefact fourni : export.csv, 4000 lignes
Critères de succès : chaque ligne correspond à une commande existante dans la base
Outils fournis : lecture du csv, requête en lecture sur la base
Rapport rendu : aucune ligne divergente trouvée, 4000 lignes contrôlées une par une

The last field in this example, four thousand lines checked one by one, is what makes the empty report credible. Without this count, the same sentence would prove nothing more than a check that never ran.

A panel that compares, a critic that looks for what is missing

A panel of judges compares several solutions proposed for the same task and scores them against each other. This is a widespread practice in the language model evaluation field, but neither of the two official pages devoted to multi-agent systems names this pattern: the vocabulary comes from common usage, not from Anthropic.

A completeness critic plays a different role, documented under this exact name: an agent whose only job is to compare what the brief asked for against what was actually delivered, to flag what is missing. It never rereads content already present looking for an error, that work belongs to the verifier subagent or to the adversarial review described in an unverified premise in a prompt comes back as a conclusion.

Mixing these three roles into a single pipeline dilutes each of them. A verifier also asked to compare several versions loses its useful bias towards rejection, a completeness critic asked to reread the substance loses the one advantage of its narrow scope.