The numbers tell a story that’s hard to ignore. When you open OpenRouter’s weekly application rankings, the CLI Agent category shows a clear front-runner: Hermes Agent with 4.91 trillion tokens consumed, nearly four times more than the second-place OpenClaw at 1.25 trillion tokens. This gap isn’t just a statistical blip—it represents a fundamental shift in how the developer community is adopting AI agent frameworks.
While the Chinese market’s early enthusiasm for OpenClaw created a temporary hype bubble, Hermes Agent from Nous Research is quietly building a different kind of momentum. Launched on February 25, 2026, the open-source framework accumulated over 180,000 GitHub stars by late April, a sevenfold increase from its 27,000 stars just two months prior. No other open-source agent framework has grown faster in 2026.
The growth rate isn’t the story—it’s a symptom of deeper architectural advantages.
The confusion around Hermes is understandable. When I published the first edition of The Hermes Agent Orange Book two months ago, reader questions poured in. Some thought it was about the luxury brand Hermès. Others asked if it was another seafood-themed AI project, a reference to the "lobster" nicknames used for certain Chinese agent systems. The core question remained: what makes it different from OpenClaw?
Let’s clear the air first. Hermes is not a fashion accessory. It’s an open-source AI agent framework developed by Nous Research, available on GitHub under the repository NousResearch/hermes-agent. If you’ve been using productized agents like OpenClaw, you might recognize the use cases: set it to scrape hundreds of AI-related X posts each morning, curate the most relevant ones, and deliver them to your inbox. Or have it monitor three competitor websites simultaneously, pinging you via Telegram whenever a new feature launches—while you sleep. It can book flights, organize your inbox, run market research, and write weekly reports.
But these capabilities aren’t what separate Hermes from the pack. Every modern agent can do these things with enough configuration. The true differentiator lies in Hermes’ ability to learn autonomously, without requiring you to become its personal trainer.
Consider the typical agent onboarding experience. With OpenClaw, you must write a detailed "soul.md" file that dictates your agent’s behavior and personality. Without this manual prompt engineering, the agent remains an empty shell—polite but useless. You invest significant time upfront to shape its responses, and every adjustment requires another round of explicit instructions.
Hermes flips this model completely. When you use it, it begins to extract patterns from your interactions. It creates compact "cheat sheets" for itself—small, reusable templates distilled from successful tasks. The next time you ask for something similar, it retrieves the relevant cheat sheet and executes without expecting you to repeat yourself. When things go wrong, Hermes observes your feedback and modifies its cheat sheets accordingly.
This isn’t theoretical. The framework employs an internal "housekeeper" process that periodically reviews all accumulated cheat sheets, merging duplicates, archiving outdated patterns, and flagging failed approaches. Combined with an automatic task persistence system that keeps goals alive even when interrupted, these three components form what the developer community calls "self-growing reins." The agent doesn’t need you to train it—it trains itself.
The architectural philosophy here mirrors what I outlined earlier this year in The Orange Book of Harness Engineering, available on WeChat Reading. That book argued that building effective agents requires five components: instructions, constraints, feedback, memory, and orchestration. The subtext was clear: agents are powerful, but you must build a custom control system to use them well. Hermes took this framework and baked it directly into its baseline, automating the harness engineering process that users previously had to manage manually.
The fundamental distinction between Hermes and OpenClaw is the difference between adopting a self-reliant colleague and training a demanding pet.
Why did I completely rewrite the book between April and now? The first edition was based on version 0.7.0. Hermes has since iterated through nine major revisions to reach 0.16.0. The framework has transformed significantly. Key improvements include enhanced memory compression, more robust error recovery, and a streamlined migration command, claw migrate, that imports your existing OpenClaw configuration in one line. This isn’t just an upgrade—it’s a product that has outgrown its original documentation.
To put this in broader context, the AI agent landscape is currently at an inflection point. Early agents like AutoGPT relied on naive iteration with no persistent learning. More recent frameworks like Crew AI emphasized multi-agent orchestration but required substantial manual tuning. Hermes represents a third generation where the agent’s behavior evolves organically, reducing the burden on users while increasing performance over time.
Critics might argue that self-learning agents risk becoming black boxes, where users lose visibility into the logic driving decisions. This concern has merit, especially in regulated industries. However, Hermes addresses this by maintaining transparent cheat sheets that users can inspect and edit if desired. Learning doesn’t have to mean opaque—it can mean verifiable growth.
The OpenRouter token consumption data reveals something else: Hermes’ lead isn’t just a developer novelty. With 4.91 trillion tokens consumed, it’s being used at industrial scale. By contrast, OpenClaw’s 1.25 trillion tokens, while impressive, suggests more fragmented or smaller-scale adoption. This gap likely reflects enterprise and power users who see long-term value in a framework that rewards continued use.
For those still on the fence, consider this: every hour you spend configuring OpenClaw’s soul.md is an hour you could have spent letting Hermes learn your preferences without explicit instructions. The early setup cost is lower, and the return on time invested compounds faster.
The question isn’t whether Hermes will surpass OpenClaw in adoption—it’s whether you can afford to keep manual-tuning your agent while competitors hire a self-learning digital worker.
The most important takeaway here is practical. You don’t need to understand the internal architecture to benefit. Download the framework, test it on a mundane task like email summarization for three consecutive days, and observe how its summaries improve without any changes on your part. That’s the moment the concept becomes tangible.
The market has spoken through numbers: 180,000 GitHub stars, 4.91 trillion tokens, and growing fast. This isn’t hype—it’s adoption driven by a genuinely superior product philosophy. Hermes isn’t asking you to learn how to tame it. It’s asking you to let it learn how to serve you.