BOLT Agent Control Plane

BOLT Agent Control Plane explores a control layer between AI reasoning and consequential actions. The prototype classifies planned actions as allow, approval_required, or deny. NVIDIA Nemotron-3-Nano-30B-A3B runs on Nebius Token Factory as the reasoning layer for a governed workflow. In a validated live run, internal project summarization and email drafting were classified as low-risk allowed steps, while sending the email to an external partner required human approval. The public repository contains a Node.js policy harness, automated policy tests, an OpenAI-compatible Nebius Token Factory integration, a hosted interactive demo, and captured evidence from the validated Nemotron run. The model was exercised directly in Nebius Token Factory Playground. The observed run reached first token in about 168 ms, completed in about 2.9 seconds, and reported roughly 183.2 tokens per second. The central challenge is separating useful model reasoning from execution authority: model output proposes a plan, but the final action passes through an explicit BOLT policy boundary. Credentials are kept out of the repository and the public demo is deterministic. Next steps are a larger benchmark set, approval-frequency and recovery metrics, structured execution telemetry, and evaluation of multiple Nemotron variants.