OncoAgent: A Dual-Tier Multi-Agent Framework for Privacy-Preserving Oncology Clinical Decision Support

Why does oncology need a specialized AI assistant? Oncology is one of the most information-dense and cognitively demanding fields in medicine. With guidelines from organizations like NCCN and ESMO evolving rapidly, it's nearly impossible for clinicians to keep up with every piece of evidence. Existing general-purpose medical AI systems often fall short due to three key issues: hallucinated recommendations, privacy risks from cloud API dependency, and monolithic models struggling with complex, multi-comorbidity cases. OncoAgent was built specifically to tackle these challenges—not to be a generalist, but to solve concrete problems in cancer care decision-making. How does OncoAgent work? The core philosophy is “divide and conquer” with multiple layers of validation. First, there’s a dual-tier model architecture. The system automatically scores the complexity of each clinical query: straightforward questions (like drug dosage checks) are routed to a lightweight 9B-parameter model for quick responses, while complex cases (like treatment decisions after multiple lines of therapy) go to a 27B-parameter model for deeper reasoning. Think of it as a triage system for AI—simple cases get fast answers, complex ones get thorough analysis, optimizing both speed and accuracy. Second, it employs a four-stage Corrective RAG (CRAG) pipeline. This isn’t just “retrieve and generate.” The system goes through retrieve → generate → evaluate → refine. It pulls information from over 70 authoritative guidelines (NCCN, ESMO, etc.), generates an initial recommendation, then has a dedicated evaluation node check its relevance to the source material. If relevance is low, it triggers re-retrieval or refinement, ensuring the final output is firmly grounded in evidence—dramatically reducing the risk of hallucinations. Third, multi-agent collaboration via LangGraph. The entire clinical reasoning process is broken down into 8 specialized nodes (agents) within a LangGraph framework—each with a single, auditable function (e.g., query understanding, retrieval, safety review). This turns decision-making into a transparent, step-by-step pipeline. If something goes wrong, you can pinpoint exactly where, which builds trust with clinicians. Finally—and most critically—privacy and hardware sovereignty. OncoAgent is fully open-source and designed for on-premises deployment. It runs on AMD Instinct MI300X GPUs, leveraging their 192GB HBM3 memory to fine-tune on 266,000+ oncology cases in about 50 minutes. More importantly, it enforces a strict Zero-PHI (Protected Health Information) policy. All data processing happens locally within the hospital’s infrastructure, with no reliance on external cloud APIs, directly addressing the core conflict in healthcare: data can’t leave the premises. Broader trends: What does this reveal about AI in healthcare?

1. From monolithic LLMs to agent-based systems: Standalone models hit limits in high-stakes scenarios. The future lies in multi-agent systems where specialized agents collaborate through orchestrated workflows—exactly the approach OncoAgent demonstrates. 2. “Hardware sovereignty” as a necessity: In sectors like finance, healthcare, and government, “data stays local” is non-negotiable. Solutions that enable full-stack, on-prem deployment using high-performance local hardware (like AMD MI300X) will have a major edge. This isn’t just a software challenge—it’s about building integrated hardware-software ecosystems. 3. Deep vertical specialization: General-purpose models are often “jack of all trades, master of none” in professional domains. True value comes from deep customization—using domain-specific data (266K cases) for QLoRA fine-tuning and tightly integrating domain knowledge bases (CRAG). This verticalization will be key for AI in law, finance, and other specialized fields. Practical takeaways and a counterintuitive insight For developers and healthcare IT professionals, OncoAgent is a valuable open-source blueprint. It showcases how to orchestrate multi-agent systems with LangGraph, design CRAG pipelines, and perform efficient fine-tuning under resource constraints (QLoRA + sequence packing). Even outside healthcare, its architectural principles are transferable to any domain requiring high reliability and strong privacy. A counterintuitive point often overlooked: pursuing “absolute safety” (like zero privacy leakage) and “peak performance” aren’t at odds—they can be achieved together through clever architectural design (like dual-tier routing). OncoAgent doesn’t sacrifice intelligence for privacy, nor does it compromise privacy for capability. It proves you can build powerful AI systems within strict constraints, offering a roadmap and confidence for anyone deploying AI in highly regulated industries.