Skip to content
Hacker News front page

How Bayer built a reliable multi-agent RAG system for preclinical research

Building Reliable Agentic AI Systems

Bayer and Thoughtworks built PRINCE, a platform that uses specialized agents—intent clarification, research, reflection, and writing—plus RAG to help scientists query decades of safety reports and draft regulatory documents. Reliability comes from three design choices: full traceability per step, continuous evaluation against test sets, and human-in-the-loop at critical checkpoints. The post doesn't disclose accuracy metrics or latency figures, but it details how the reflection agent checks data sufficiency and answer grounding. The architecture is solid, though the operational overhead is non-trivial—best suited for teams with strict compliance needs.

Why it matters: Bayer and Thoughtworks split retrieval, checking and drafting across four agents, a reusable pattern for making multi-agent systems dependable.

Read the original ↗Export Markdown