| I built SteerPlane — an open-source runtime guardrail system for AI agents. The problem: AI agents run autonomously, calling LLMs and tools in loops. Without runtime controls, a single agent can burn $50+ in minutes, get stuck in infinite loops, or call dangerous actions without oversight. What it does:
- Cost ceilings — auto-kills when spending exceeds $X
- Step limits — prevents runaway execution
- Loop detection — catches repeated action patterns
- Full telemetry — every step logged with tokens, cost, latency
- Dashboard — real-time visibility into all agent runs One decorator. That's it:
@guard(max_cost_usd=10, max_steps=50)
def run_agent():
agent.run() Stack: Python SDK, TypeScript SDK, FastAPI backend, Next.js dashboard Links:
- GitHub: https://github.com/vijaym2k6/SteerPlane
- PyPI: pip install steerplane
- npm: npm install steerplane Currently building: policy engine (allow/deny actions), remote kill switch, and framework integrations. Would love feedback from anyone running AI agents in production! What controls do you wish you had? |