Here's why it matters: AI agents are starting to take real actions in production — moving money, modifying records, talking to customers. Traditional observability tools were not designed for this. They tell you what happened after the agent acted. By the time you've read the log, the refund cleared, the record was deleted, the email was sent twelve times.
While many observability vendors have tried to extend into agent workloads, the engineers we've talked to keep asking for something different: a layer that sits inline, before tool execution, and prevents bad actions instead of just logging them.
SafeRun is built from the ground up for that. Inline policy evaluation. Loop and cost circuit breakers. Human-in-the-loop approval queues. Frame-by-frame replay debugging.
We're building this as developer infrastructure — a Python or TypeScript decorator that wraps tool execution in three lines of code, with native integrations for LangGraph, OpenAI Agents SDK, Anthropic Claude Agent SDK, Vercel AI SDK, CrewAI, and Mastra. Or sit at the MCP layer for framework-agnostic coverage.
One thing is certain: AI agents are taking real actions, and you can't ship them to production without a layer that says no.
With SafeRun, teams can:
→ Validate every tool call against declarative policies before execution
→ Break runaway loops and circuit-break on cost overruns
→ Escalate ambiguous actions to a human approval queue via Slack
→ Replay every agent decision frame by frame when something breaks
→ Deploy as managed SaaS or self-hosted in your VPC
Opening up early access soon — let me know if you're interested in the comments or a DM.
saferun.dev
Top comments (0)