Before a brochure or a slide deck reaches a clinician, someone has to trace every clinical claim in it back to its source. Medical, legal and regulatory reviewers do this by hand, moving between study PDFs, FDA labels, PubMed abstracts and ClinicalTrials.gov records to verify a single page, and an asset takes around 15 days on average to clear. The downside of getting it wrong isn't theoretical: the FDA issued more than 100 cease-and-desist letters for misleading promotional content in September 2025 alone.

Medical Content Review is an open-source application on Amazon Bedrock AgentCore that automates the first pass. You upload the content with its supporting references, and a multi-agent system reads every page and returns findings with severity scores, suggested fixes and citations.

The design point worth dwelling on is that this workload is not open-ended research. The agent reads every page and checks every claim, skips nothing, and never decides for itself what looks interesting, so the orchestrator fixes the order. It runs OCR on each PDF through a multimodal model, groups pages into review units, then fans out three specialists per unit. Editorial checks language and figure consistency and carries no external tools, internal references compares claims against the uploaded documents, which is the only way to cover pre-launch drugs, and external evidence queries PubMed, OpenFDA and ClinicalTrials.gov. It's a foundation rather than a finished MLR system, and it builds on the AgentCore Deep Research sample.