The Isolated Reviewer Pass: why critique can't share context with the drafter

Why RoleSherpa's reviewer agent never reads the drafter's own reasoning — only its output.

Part of the Keeping AI agents in check case study.

The failure mode this prevents

The easy way to build a "review" step into an agentic pipeline is to hand the same conversation — drafter's reasoning included — to a second call and ask it to critique the result. It's cheap to build and it produces plausible-looking output. It also tends to agree with itself: a model that can see why the drafter made a choice is primed to accept that reasoning rather than independently evaluate the result against the actual requirement.

That failure mode is easy to miss in a demo, because the reviewer still produces some commentary — it's just commentary that rarely catches the kind of mistake the drafter itself would have caught.

The isolation, specifically

RoleSherpa's reviewer runs as a separate agent dispatch: it receives the drafted CV/cover-letter output and the job posting, but not the drafter's intermediate reasoning, tool calls, or context window. It researches the target company independently and critiques the artifact cold — the way an actual second opinion would.

Verification is a separate, later gate

The isolated reviewer pass is not the same thing as the anti-fabrication verification step described in the Keeping AI agents in check overview — the two are deliberately separate gates. The reviewer pass is a qualitative critique (does this read well, is it well-targeted to the company). The verification step that follows is a mechanical check (page count, real text-layer extraction, keyword coverage, and a claim-by-claim diff against the stored profile). A document has to pass both, independently, before it's shown to the user.

Why this is worth the extra dispatch

Running an isolated reviewer costs an additional agent call per application — RoleSherpa spends that cost deliberately, because the alternative (a reviewer that shares the drafter's context) produces a false sense of having been checked, which is worse than not reviewing at all.