Runtime guardrails as a principle for agent design.

[ishita-0301]

One principle I'd add to the "12 factors" for building reliable agents is fail safely at runtime.
In practice, I've seen more issues from execution than from prompting. Agents get stuck repeating the same tool call, endlessly retry a failed action, or keep consuming tokens without making progress.

I've been experimenting with an open source project called FailproofAI that focuses on this exact problem. Instead of changing the model, it adds runtime guardrails like loop detection and execution policies to help agents recover gracefully.
Repository: https://github.com/FailproofAI/failproofai

Curious if others think runtime reliability deserves to be a first class concern alongside statelessness, observability, and context management when designing production agents.