Nous Research dropped Hermes Agent v0.18.0 today, and it’s not just another version bump. They call it the “Judgement Release,” and the headline is the kind of claim most open source projects never make: every single P0 and P1 issue in the entire repository is closed. Zero remain open.
That’s roughly 692 highest-priority items resolved over twelve days, cleaning up 496 P0/P1 issues and 196 matching PRs. The team apparently pulled all-nighters right up to the release cut to close the final cluster, an interrupt-protected compression sibling-fork bug and its fix. And they intend to keep P0/P1 at zero from here forward.
But the bug sweep is just the frame. The real story is what v0.18.0 actually ships.
Mixture-of-Agents Is Now a First-Class Model
The MoA feature that rolled out in preview back in June is now a full selectable model type. Instead of toggling a mode, you pick a named MoA ensemble like “my-council” in every model picker (CLI, TUI, desktop app, gateway), right alongside Claude, GPT, and Grok.
Every reference model in the ensemble now shows its full reasoning in its own labelled block. You read what GPT-5 thought, what Claude thought, what Grok thought, then watch the aggregator synthesize them into an answer that streams live. No more staring at a blank screen while the committee deliberates.
Hermes Agent Actually Verifies Its Own Work
This one matters. “Done” in Hermes now means the agent ran your project’s actual checks, not that it feels good about finishing. The verification engine records evidence for coding work, and the /goal command now supports completion contracts. You state what “done” looks like, and the standing-goal loop judges completion against real evidence instead of stopping when the model decides it’s satisfied.
There’s a pre_verify hook for custom checks, and a one-time migration tunes the defaults sensibly. The difference between “I think I fixed it” and “the tests pass, here’s proof.”
/learn and /journey Turn Self-Improvement Into Something You Can See
The /learn command that launched in late June is now fully baked in Hermes Agent: point it at a directory, a URL, or just the workflow you walked through five minutes ago, and it distills a reusable skill that follows your CONTRIBUTING.md conventions automatically. Teaching Hermes a new trick is one command.
/journey is new in v0.18.0. It’s a playable timeline of everything Hermes has learned about you (memories and skills the agent accumulated over time), and you can edit or delete any of them right from the view. Paired with the desktop app’s new memory graph (a radial timeline), your agent’s memory stops being a black box.
Background Subagent Fan-Out
delegate_task can now dispatch multiple subagents that all run in the background. Your chat never blocks, and when every subagent finishes, their results consolidate into a single turn. Kick off “research these five competitors in parallel” or “audit these three modules,” then carry on with something else while a small fleet works. When it’s all done, you get one clean summary instead of babysitting each one.
Desktop Coding Projects
The desktop app in Hermes Agent gained real, per-profile Projects (a sidebar of your codebases, a coding rail, a review pane, git worktree management, and agent-facing project tools), all backed by a proper project-into-repo-into-lane model. Instead of scattered chat sessions, your coding work is organized into projects the agent understands and can act on. It’s the desktop turning into an actual coding cockpit.
Production-Ready Gateway Infrastructure
For anyone running Hermes for a team or as a hosted service, the gateway can now scale to zero when idle and quiesce cleanly before restarts, migrations, or auto-updates without dropping in-flight conversations. Disruptive lifecycle actions coordinate an external drain so nobody gets cut off mid-turn.
There’s also a new /prompt command that opens your $EDITOR for multi-line markdown prompts instead of fighting a one-line input box, and cheaper self-improvement via an auxiliary model that digests context instead of replaying the whole conversation.
Google Vertex AI is now a first-class provider for Gemini models, handling the short-lived OAuth2 token minting and auto-refresh automatically, a huge quality-of-life fix for anyone running Hermes through a GCP service account.
Bottom Line
v0.18.0 is the most consequential Hermes Agent release in months. The P0/P1 clean sweep is a genuine engineering flex, but the new capabilities (first-class MoA, self-verification, /learn and /journey, background fan-out, desktop projects, production gateway scaling) move the agent from “powerful CLI tool” toward “platform you can actually build workflows on.”
It’s available now on GitHub. If you’re on an earlier version, the changelog from v0.17.0 alone runs ~1,720 commits and 998 merged PRs. It’s a big jump.
