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

OncoAgent Framework Released for On-Premises Clinical AI

A research group has released OncoAgent, an open-source clinical decision support system designed specifically for oncology. The framework aims to solve persistent issues in medical AI by combining a multi-agent architecture with a strict, privacy-preserving design that allows it to run entirely within a hospital's local infrastructure. By avoiding dependency on external cloud APIs, the system directly addresses the critical data sovereignty requirements of healthcare institutions and mitigates risks associated with model hallucinations in high-stakes clinical settings.

System Architecture and Performance

The system's core is a dual-tier model architecture built on a LangGraph state machine. Clinical queries are first scored for complexity and then routed to either a 9B parameter model for rapid triage or a 27B model for deep reasoning. Both models were fine-tuned using QLoRA on a dataset of over 266,000 oncological cases. All model outputs are grounded against a knowledge base of over 70 physician-grade guidelines from NCCN and ESMO using a four-stage Corrective RAG (CRAG) pipeline. The entire training process was optimized with the Unsloth framework on AMD Instinct MI300X hardware, which enabled a 56x throughput acceleration, completing a full fine-tuning run in approximately 50 minutes.

• Framework: LangGraph multi-agent topology
• Models: Dual-tier Qwen 9B & 27B fine-tuned with QLoRA
• Knowledge Base: 70+ NCCN & ESMO oncology guidelines
• Retrieval: Four-stage Corrective RAG pipeline with relevance grading
• Hardware: Trained and deployed on AMD Instinct MI300X GPUs
• Policy: Strict Zero-PHI for on-premises deployment

Implications for Clinical AI

OncoAgent represents a significant architectural pattern for AI in regulated industries. Its open-source, on-premises approach provides a viable alternative to monolithic, proprietary models that often present data privacy and auditability challenges. By decomposing clinical reasoning into auditable steps and enforcing a mandatory human-in-the-loop gate for complex cases, the framework provides a blueprint for building trustworthy AI systems. This push for "hardware sovereignty" enables medical centers to leverage advanced AI capabilities without compromising control over sensitive patient data, potentially accelerating adoption in privacy-conscious environments.

OncoAgent's success is less about a single model and more about a strategic integration of open-source components—LangGraph for logic, Unsloth for hardware optimization, and QLoRA for efficient tuning. This stack demonstrates a clear path for enterprise AI that prioritizes data sovereignty and vertical-specific performance over reliance on closed, general-purpose cloud APIs, a critical pattern for regulated fields like healthcare and finance.