Hermes Agent v0.18.0 Closes Every Critical Bug and Learns to Prove Its Work

Nous Research shipped Hermes Agent v0.18.0 this week, and the headline isn’t a feature — it’s housekeeping at a scale you almost never see in open source. The team spent a week and a half closing every single P0 and P1 issue in the entire repository. Roughly 700 highest-priority items, out of nearly 2,000 total issues and PRs closed in the window. Zero remain open, and they say they intend to keep it that way.

If you’re not a developer: P0s and P1s are the “drop everything” bugs. Most projects carry a permanent backlog of them. Clearing the board completely is the software equivalent of a restaurant closing for a week to deep-clean the kitchen — and then promising health-inspector-ready every day after.

But the cleanup is the frame, not the story. What actually shipped matters more.

You can now pick a committee of AIs like it’s one model

Mixture-of-Agents — asking several AI models the same question and having one synthesize the answers — used to be a mode you toggled. Now a named ensemble shows up in every model picker like any single model. Pick “my-council” the way you’d pick Claude or GPT, and you can watch each model’s reasoning in its own labeled block before the combined answer streams in. For hard questions, a panel beats a single expert — and it’s now one selection away.

”Done” now means proven

The feature that matters most if you actually let agents work unsupervised: Hermes now verifies its own work. For coding tasks, it runs your project’s real checks and records the evidence, instead of declaring victory when the model feels finished. The /goal command adds completion contracts — you define what “done” looks like, and the agent judges itself against that, not its own optimism. That’s the difference between “I think I fixed it” and “the tests pass, here’s proof.”

Its memory stops being a black box

Two new commands round out the self-improvement loop. /learn distills anything — a folder, a URL, a workflow you just walked it through — into a reusable skill. /journey shows a timeline of everything the agent has learned about you, memories and skills included, and lets you edit or delete any of it. If the idea of an AI quietly building a model of you makes you uneasy, this is the right answer: make it visible and editable.

Also in this release

Background subagent fan-out (delegate several jobs, keep chatting, get one consolidated summary), real coding Projects in the desktop app, gateway scale-to-zero for anyone hosting Hermes for a team, and Google Vertex AI as a first-class provider.

Bottom Line

This is the most consequential Hermes release in months, and the pattern is worth noticing beyond this one tool: the agent world is shifting from “look what it can do” to “look how you can trust it.” Self-verification, visible memory, and a zero-critical-bug pledge are all the same move. If you run Hermes, update. If you’re just agent-curious, this is what maturing looks like.