Hermes agent loops point to a need for safety limits

can sometimes repeat the same action or keep trying the same tool call many times. That can burn through tokens quickly and use up a daily budget within a few hours.

This makes fully unsupervised runs hard to trust, because a person may need to watch the agent and stop wasteful behavior. One proposed fix is a that detects loops, stops them, repairs the problem during the run, and saves the fix in a database so later sessions can avoid the same failure.

A shared knowledge database could also let many people collect fixes for common agent mistakes. The open questions are whether this is mainly a usage problem, a common Hermes behavior, or something that improves with a different model.

Key points

  • Repeated actions and repeated tool call attempts can waste tokens fast.
  • Unsupervised Hermes runs may need budget and loop limits before they are safe to trust.
  • A could detect loops, stop them, fix the issue, and remember the solution.
  • Saving fixes in a database may help future sessions avoid the same failure.
  • A shared knowledge database could collect common fixes, but it needs validation.
Read original