Home / Several agents and adversarial checking

An unverified premise in a prompt comes back as a conclusion

The framing text an orchestrator writes into each agent's prompt stays invisible to adversarial review, which receives the produced artefact and the success criteria, never the intent or the history of whoever built the task: a false claim slipped into that framing therefore travels through the whole chain untested, and a fan-out that sends the same framing to several agents multiplies this exposure instead of diluting it.

An adversarial verifier receives the artefact to be judged and the success criteria it must test, never the history of the task nor the reasoning of the person who built it, according to the rationale Anthropic documents for this role. This deliberate limit, which keeps the verifier focused on what is in front of it, leaves one precise blind spot: it says nothing about what the orchestrator had already assumed to be true before even writing the first prompt.

The framing travels without ever being tested

The introductory text of a prompt often states a factual claim presented as already established rather than as a conclusion to be checked, for example that supplied data is already clean or that a scope has already been validated by a third party. The agent that receives this prompt treats that sentence as a premise, not as a fact to challenge, and reflects it back in its own output without ever restating it as a hypothesis. The adversarial verifier, for its part, only reviews that output: the original premise appears nowhere in a form it can put to the test.

Prompt sent to each agent in the fan-out:
"The catalogue provided contains only active references.
Write the product sheet for each of the listed references."

The adversarial verifier receives only the sheets produced,
never the sentence about active references.

The fan-out multiplies the exposure instead of diluting it

A fan-out sends the same framing text to every agent launched in parallel. A false premise slipped into that text is therefore not diluted by the number of agents, it is repeated in each of their outputs. Several independent agents that separately arrive at the same claim can then give the impression of corroboration, when in fact they all share the same source, the same starting prompt, and the same blind spot. Assigning a reviewer to specifically challenge the framing, not just the conclusions built on top of it, gives adversarial review a grip on this sentence that it normally never sees. The previous lesson shows what an adversarial verifier receives to judge a piece of work; here, the problem comes precisely from what it never receives.

Sentence in the prompt: the catalogue provided contains
only active references.
Never tested, never shown to the verifier.

Product sheet: active reference, in
stock.
Cited as an established fact, without ever being verified.