Runtime failures in manager-worker setups

[ishita-0301]

Looking at HiClaw's architecture, one thing I kept wondering about is what happens when a worker silently gets stuck.

It's easy to imagine a worker repeatedly trying the same thing, burning tokens, while the manager assumes progress is being made. Feels like there could be value in surfacing signals like "no forward progress" or "repeated identical actions" at the orchestration layer.

I've seen similar ideas discussed in projects like FailproofAI (https://github.com/FailproofAI/failproofai), where the focus is on detecting runtime failures rather than improving the model itself. Curious if anyone building multi-agent systems has run into this in practice.