SteerPlane
Runtime Control Plane for Autonomous AI Agents - cost limits, loop detection, and full observability with one decorator
A runtime guardrail library that wraps autonomous AI agents with cost limits, loop detection, and observability using a single decorator.
SteerPlane is a runtime safety layer for autonomous AI agents, the kind of software that can call APIs, run code, and take actions on its own. Without any oversight, that kind of agent can get stuck repeating the same actions in a loop, spend an enormous amount of money on API calls overnight, or take a destructive action with nobody noticing until it is too late. SteerPlane's stated goal is to prevent those failure modes by wrapping an agent's run loop with limits and monitoring, added with a single decorator around the function that runs the agent.
Once wrapped, every step the agent takes is logged along with its token usage, dollar cost, and how long it took. You can set a maximum dollar budget and a maximum number of steps for a single run, deny specific kinds of actions by name pattern, such as anything starting with delete or sudo, and set rate limits on how often a particular action can happen in a given time window. SteerPlane can either kill a run outright when a limit is broken, or pause it and alert a human for approval before continuing, depending on how it is configured. It also includes a sliding window algorithm for detecting when an agent is repeating the same action or pattern of actions, catching this without needing an extra call to a language model to check.
Beyond the core Python and TypeScript libraries, the project includes a proxy gateway that can enforce these same limits on API traffic without changing any application code, a real time web dashboard for watching runs as they happen, a command line tool for managing runs and API keys, and ready made integrations for popular agent frameworks including LangChain, the OpenAI Agents SDK, CrewAI, and AutoGen. If the central service is ever unreachable, the underlying library keeps enforcing limits locally so an agent is never left unprotected.
The core project runs on Docker with a bundled database, and its code is released under the MIT license, though some advanced features such as human approval notifications are reserved for a separate hosted plan.
Where it fits
- Cap how much an autonomous AI agent can spend before it gets shut down.
- Detect and stop an AI agent stuck in a repeating action loop.
- Add a human approval step before an AI agent can exceed a cost or action limit.