🤚 The Open-Palm Illumination

The open-source AI agent wars have produced their first funeral, and the eulogy is being delivered in a terminal window. Nous Research released Hermes Agent v0.18.0 — codenamed “The Judgment Release” — on July 1, 2026, and the sheer violence of the changelog has prompted at least one prominent AI creator to declare OpenClaw officially deceased.

The numbers alone read like a startup’s entire engineering history compressed into a single sprint: ~1,720 commits, 998 merged pull requests, 2,215 files changed, approximately 251,000 insertions, and contributions from 370+ community developers. The team didn’t just ship features — they cleared every single P0 and P1 issue in the repository. All ~700 highest-priority items. Closed. Done. The kind of housekeeping that makes project managers weep into their Gantt charts.

But the headline feature? Mixture of Agents (MoA) is now a first-class citizen. Not a plugin. Not a workaround. Not something you wire together with duct tape and YAML prayers. A proper, selectable model type sitting right alongside Claude, GPT, and Grok in every picker — CLI, TUI, desktop app, gateway. You pick “my-council” like you’d pick any model, and Hermes routes your prompt through an ensemble of models automatically.

👐 The Two-Handed Reality Check

Let us be specific about what Mixture of Agents actually delivers, because the concept has been floating around AI Twitter for two years like a conference talk that never shipped. In Hermes v0.18.0, when you invoke a MoA ensemble, each reference model’s full output renders as its own labelled block. You can read what Claude thought. What GPT suggested. What Grok hallucinated. And then the aggregator synthesizes them into one answer, streamed live, rather than appearing all at once like a magic trick with no showmanship.

This matters because it transforms MoA from a black box into a transparent deliberation chamber. You’re not just getting “the answer” — you’re watching multiple frontier models argue with each other in real time and then produce a consensus. It’s like hiring three expensive consultants and actually making them show their work.

Then there’s /learn — a command that lets you teach Hermes a reusable skill from essentially anything. Point it at a directory, a URL, a conversation, or paste in some notes, and the agent will source the material using its own tools and write a standards-compliant SKILL.md file. No hand-writing required. Your workflows become slash commands without you ever touching a markup file. The implications for enterprise adoption are, frankly, the kind of thing that makes product managers at competing projects stare at the ceiling at 2 AM.

And Journey — a visualization system for agent workflows — rounds out the major additions, though cost savings may be the quietest killer feature. Progressive disclosure now keeps skills cheap by loading a ~3,000-token index first, with full content loading only when a task actually needs it. Translation: your API bill just got meaningfully smaller for doing the same work.

🌿 The Gentle Awakening

The framing of “Hermes ended OpenClaw” is, of course, the YouTube equivalent of declaring a winner in a bar fight where both participants are still standing and one of them hasn’t noticed the chair yet. OpenClaw — the open-source AI agent that surpassed 100,000 GitHub stars and became the default recommendation in every “build your own AI assistant” tutorial — is very much alive. It has a massive community, deep integrations, and the kind of brand recognition that survives individual feature gaps.

But here’s the uncomfortable truth: Hermes Agent is shipping at a pace and scope that makes the comparison interesting rather than absurd. A native desktop app built in a single week across 100 PRs and 159 commits — for macOS, Linux, and Windows — with one-click install and in-app self-update. The /goal command now supports completion contracts where you state what “done” looks like, and Hermes decides it’s finished by actually running your project’s checks. Verification evidence for coding work. The agent doesn’t just write code — it proves the code works.

The philosophical gap between these projects is widening in a way that matters. OpenClaw remains the Swiss Army knife — model-agnostic, privacy-focused, extensible in every direction. Hermes is becoming something more opinionated: an agent that learns, that judges its own output, that treats multiple models as a jury rather than a menu. Whether that opinionation is a feature or a limitation depends entirely on whether you trust the kitchen.

👑 The Crown Verdict

We are witnessing the moment where AI agents stop competing on “can it do the thing” and start competing on how elegantly it orchestrates doing the thing. Mixture of Agents as a first-class feature isn’t just a technical improvement — it’s a philosophical statement that no single model is the answer, and the real intelligence lives in the ensemble. Nous Research just built a parliament where Claude, GPT, and Grok sit in session, and the aggregator is the speaker of the house.

The /learn command is equally significant, though it’ll get less attention because “the agent can now teach itself new skills from any source” sounds like a sentence from a sci-fi novel rather than a changelog. But reusable, self-authored skills are the foundation of compound intelligence — agents that don’t just execute tasks but accumulate capability over time. That’s not a feature update. That’s an evolutionary pressure.

Is OpenClaw dead? No. Is the gap between “open-source agent that does things” and “open-source agent that orchestrates, learns, verifies, and improves” getting wider? The 1,720 commits in this release would suggest yes. And in the AI agent race, the distance between “alive” and “relevant” is measured in release cycles, not GitHub stars.

Inspired by The new Hermes Agent update officially ended OpenClaw by Alex Finn.

Your ensemble is showing. Aggregate wisely.