Blue River SAFR: Governed AI Agents on AMD

Blue River SAFR is a runtime governance and assurance layer for autonomous AI agents operating in high-trust and regulated environments. Instead of allowing an agent to move directly from reasoning to execution, SAFR places a control layer between the model and the action. For each proposed action, the system evaluates agent identity, permissions, policy rules, risk, spending limits, approval requirements and evidence obligations. The action can then be approved, escalated for human review, modified or refused. Every decision produces an inspectable evidence trail showing what the agent intended to do, which controls were applied and why the final decision was made. For the AMD AI Academy Challenge, the prototype will run open-model inference and evaluation workloads using AMD Developer Cloud and ROCm, with AMD GPU benchmarking used to measure latency and throughput. The demonstration will show governed agent workflows such as financial operations, supplier onboarding and controlled autonomous transactions. The goal is to demonstrate that powerful AI agents do not have to be black boxes: they can be fast, auditable, policy-aware and commercially deployable.