Building reliable agentic AI systems
Points and comments are a snapshot, not live.
Bayer's PRINCE platform uses agentic RAG to streamline preclinical drug data retrieval.
Bayer developed PRINCE, an agentic AI system built on Retrieval-Augmented Generation (RAG), to address fragmented preclinical data across silos. The system evolved through three phases: Search (structured metadata), Ask (natural language Q&A on reports), and Do (multi-agent task execution). Architecture uses LangGraph orchestration with Researcher, Writer, and Reflection agents, plus a dedicated Think & Plan step for process reflection. Context discipline ensures each agent receives tailored context, not a monolithic prompt. Fallbacks, retries, and observability via Langfuse and RAGAS evaluation support reliability. The system handles both structured (Athena/SQL) and unstructured (OpenSearch for vector retrieval) data.
What commenters are saying
Commenters split on whether such complex agentic architectures are warranted. Some argue decomposition into specialized agents (Researcher, Writer, Reflection) is over-engineering without evidence of benefit over simpler approaches, with one noting "an agent is just a system prompt and output contracts." Others defend decomposition, citing parallels to microservices. The thread also highlights the critical role of data quality: one commenter estimates 99% of effort goes to data preparation versus 1% to agent tuning. The relatively sparse evaluation section drew criticism given the lengthy architecture description. Several commenters distinguish between search (well-suited to LLMs) and code-generation (where quality degrades over time).