# SafeRun

> Reliability infrastructure for AI agents. Inline validation, blocking, approval, loop breaking, and replay debugging for every agent tool call.

SafeRun sits between AI agents and the tools they call — validating, approving, blocking, breaking runaway loops, and replaying risky actions. It's the inline reliability gate that keeps agentic systems from breaking production.

## Pages

- [Home](/): What SafeRun is, how it works, and the SDK.
- [How to Secure AI Agents](/how-to-secure-ai-agents): Practical guide to declarative tool-call policies, human-in-the-loop approvals, loop breakers, and replay debugging.
- [Agentic AI Risks](/agentic-ai-risks): The 8 failure modes that take agentic AI down in production and how to mitigate each.
- [AI Agent Monitoring](/ai-agent-monitoring): Signals, metrics, and tooling for monitoring AI agents in production — distinct from LLM observability.
- [SafeRun vs LangSmith](/vs/langsmith): Action control versus LLM observability — when to use each, and when to use both.