Damus
Ben Eng profile picture
Ben Eng
@Ben Eng
I am really struggling with Hermes Agent tool use (patch) to modify files when maintaining the llm-wiki. This knowledge base use case is my first experiment to see how useful and reliable Hermes can be at following instructions, executing skills, and using tools.

So far, it has been a mixed bag. I am seeing routinely edits that write intermediate representations of data as the final outcome. Sometimes line numbers are inserted to predict every line of text. Sometimes the two characters '\n' (backslash + n) are inserted along with bulk text, when clearly newline was intended. Most annoyingly, Hermes claims to make edits, indeed confidently reporting verifications using file timestamps and git status, but those results are hallucinations based on what Hermes says are 'simulations', not actual tool use.

That doesn't give me a lot of confidence to expand my use of Hermes to do more important work.

1
Ben Eng · 5w
Thanks to Gille in the Nous support channel (again!), we've concluded the blame lies with x-ai/grok-4.1-fast being quirky in tool use. Switching back to Google/gemini-3.1-flash-lite has made everything work fast and reliably again. Looks like the previous generation of Grok models is being removed M...