Hermes Agent V0.18 Judgement Release: Goal Loops and Mixture of Agents

Key Takeaways

• The “Judgement” theme is self-verification. V0.18 leans into having Hermes judge its own output quality instead of trusting a model that simply reports “done.”
• Improved /goal loop. You set a definition of done; a builder produces work, a judge scores it, and it iterates for a number of rounds you configure — running autonomously until the judge passes the quality bar.
• Proof-of-work for coding. Hermes records evidence and confirms a coding task is finished by actually running the project’s checks, closing the gap between “I think I fixed it” and “the tests actually pass.”
• Mixture of agents is now a first-class model choice. It fuses answers from multiple models (the demo pairs Claude Opus 4.8 with GPT-5.5) into one higher-quality, one-shot output — roughly 10–15 minutes, with no back-and-forth chat.
• New commands. /learn ingests a guide or repo link and saves it as a reusable skill; /journey shows a timeline of what Hermes has learned about you; background fan-out runs delegated sub-agents asynchronously so your chat isn’t blocked.
• Three ways to update. Type “update to the latest version” in chat, click Update Hermes in the dashboard, or run hermes update in the terminal.

Commands & Code Mentioned

/goal        # define a done-state; builder + judge loop until quality passes
/learn       # feed a guide or repo link; Hermes saves it as a reusable skill
/journey     # timeline of what Hermes has learned about you
hermes update  # update to the latest version from the terminal