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

Source: HuggingFace Date: 2026-05-09 URL: https://huggingface.co/blog/lablab-ai-amd-developer-hackathon/oncoagent-official-paper

Summary

OncoAgent is a hackathon project demonstrating production-grade clinical AI running entirely on-premises using open-source tooling (LangGraph, Unsloth, ROCm, ChromaDB). The system routes oncology cases through an 8-node LangGraph graph — Tier 1 (Qwen 3.5-9B) for routine cases, Tier 2 (Qwen 3.6-27B) for complex ones — with corrective RAG grounded on 70+ NCCN/ESMO guideline PDFs, a deterministic safety critic, per-patient memory isolation, and a mandatory human-in-the-loop gate for high-complexity decisions. PHI is redacted before any LLM invocation, making HIPAA de-identification structural rather than policy.

Implications

• Feeds the local/sovereign AI and multi-agent architecture threads: demonstrates that the full agentic stack (routing, retrieval, reflexion, HITL) can run on AMD MI300X hardware without cloud API calls, relevant to any operator who cannot accept data exfiltration risk.
• The deterministic critic pattern — safety validation as code, not LLM logic — is a concrete design choice that prevents adversarial prompt bypass and is directly portable to non-medical agentic systems.
• The dual-tier routing approach (complexity score gates model size) is a cost-control pattern worth tracking as agent orchestration frameworks mature; it maps cleanly onto the “right model for the task” conversation happening across the CLI agent space.