I Built a Runtime Control Plane to Stop AI Agents From Burning Money

Two weeks ago, I watched an AI agent burn through $47 in API credits in 20 minutes. It had gotten stuck in a search loop — querying the same thing over and over — and nobody noticed until the OpenAI bill came in. This wasn't a toy experiment. This was a real agent, doing real work, in a real system. And it's not just me. 87% of enterprises deploying AI agents have no security framework for their autonomous systems. Agents can call APIs, execute code, browse the web, and spend real money — with zero guardrails. So I built SteerPlane. What Is SteerPlane? SteerPlane is an open-source runtime control plane for AI agents. It sits inside your agent code (not as a proxy or gateway) and enforces safety policies in real time: Cost limits : Set a hard USD ceiling. If your agent hits it, the run is terminated instantly. Loop detection : A sliding-window algorithm watches for repeating action patterns. If your agent does [search → think → search → think] 20 times, SteerPlane catches it. Step limit