Curious how people are handling agent failures in long running workflows

[ishita-0301]

I've been using tools like Khoj for research and personal knowledge retrieval, and one thing I've started thinking about is what happens after retrieval. Once agents begin taking actions over longer sessions, they can sometimes get stuck repeating the same steps or endlessly retry a failing operation.

I recently came across an open source project called FailproofAI that focuses on these runtime issues instead of retrieval itself. It adds guardrails like loop detection and recovery to help agents make forward progress rather than burn tokens on the same mistake.

Repository: https://github.com/FailproofAI/failproofai

I'm curious whether anyone here has seen similar failure modes with autonomous workflows built on top of Khoj and what techniques you're using to handle them.