Hermes Agent vient de mettre à mal OpenClaw avec une nouvelle mise à jour folle

Hermes Agent just went from being one of those interesting open- source projects people were quietly testing to suddenly sitting at the center of one of the biggest agent stories right now. And the crazy part is that in the space of a few days, Hermes overtook OpenClaw on Open Router's global daily app and agent rankings, started processing around 224 billion tokens per day, pushed past OpenClaw's daily number, and then dropped a massive update that makes it feel like a personal AI operating system that can keep learning from your work. That alone would be enough for a big story. Yet, the timing makes it even stranger. OpenClaw was basically the agent everyone was talking about a few months ago. It had the hype, the community, the integrations, the massive skill ecosystem, and the first mover advantage. Then Hermes arrives in February, grows at a ridiculous speed, ships update after update, and by May, it's suddenly challenging the project that almost defined the whole category. And that is why this Hermes situation is getting insane. On May 10th, 2026, Hermes Agent reportedly claimed the number one spot on Open Router's global daily app and agent rankings. The public numbers showed Hermes processing around 224 billion tokens per day, while OpenClaw was sitting around 186 billion. That means Hermes wasn't just catching up in some abstract way. It was actually routing more daily inference volume than OpenClaw on one of the most visible public AI agent leaderboards. Now, that does not automatically make Hermes the best agent in the world. Open router rankings are based on tracked public usage, and token volume can be noisy, especially with agents that burn through huge context during long workflows. But in a young market, visible rankings create momentum. Developers see the number, test the tool, talk about it, and that attention can turn into even more usage. And Hermes reached that point unbelievably fast. The project launched in February 2026 from Noose Research, the same lab behind the Hermes model family, and it is now sitting at roughly 147,000 GitHub stars and more than 23,000 forks. That is wild for a project that is only a few months old. OpenClaw still has the bigger all-time footprint with more than 370,000 GitHub stars and more than 9 trillion cumulative tokens compared to Hermes at roughly 6 trillion. OpenClaw is not suddenly dead. It is still massive. The real story is velocity. OpenClaw still has the bigger historical footprint, but Hermes went from launch to the top of the daily board in under 90 days. That kind of speed usually means the product is landing exactly where developers are starting to care. And with Hermes, it seems pretty clear what they care about. Hermes is not built like a normal chatbot or even a normal coding assistant. The whole design is centered around a do, learn, improve loop. After it completes a complex task, it doesn't just throw away the experience. It reflects on what worked, extracts the useful pattern, and writes a skill file it can use again later. That sounds simple at first, although the implications are massive. Most assistants start every new job with the same general intelligence and whatever context you give them. Hermes is trying to become more specific to your work over time. If it solves a repeated coding workflow or business process, it can turn that experience into reusable procedural knowledge instead of starting from zero every time. That is the difference between an AI tool that responds and an AI system that compounds. And this is exactly where Hermes is being compared to OpenClaw. Openclaw won attention by being everywhere. It connected agents to Telegram, Discord, Slack, WhatsApp, Signal, and a long list of other platforms. Hermes is making a different bet. Instead of focusing first on reach, it focuses on learning from your actual work. Openclaw has a massive skill ecosystem, but Hermes is trying to make skill creation part of the agents natural workflow. Now, right in the middle of all this Hermes momentum, I have to mention that today's video is sponsored by Higsfield because what they just launched fits perfectly into this whole agent story. Higsfield MCP is basically the first time Claude, OpenClaw, and Hermes start feeling connected to a real media production pipeline instead of just being smart tech systems. The easiest way to explain it is this. Claude gives the reasoning while Higsfield MCP gives the agent actual creative execution. So instead of Claude only talking about ads, videos, thumbnails, landing pages, or content strategy, the agent can now actually generate and ship those assets directly inside the workflow. And Hermes already supports MCP servers natively. So setup is simple. You drop the Higsfield MCP endpoint into your Hermes config. Reload the agent and suddenly Hermes can start generating videos, images, ads, landing pages, and other creative assets directly inside the pipeline. One example they showed was Hermes orchestrating multiple clawed agents through Higsfield MCP where one agent handles statics. Another handles UGC style videos, another handles email outreach, and another watches analytics and rewrites underperforming creatives automatically. So yeah, if you want to test this yourself, check out Higsfield MCP through the link in the description. All right, now back to the video. Open clause skills are often more like runbooks created upfront by a person or by an AI you prompt to write them. Hermes tries to build that into the agent itself. It works, reflects, writes a reusable markdown skill and refineses it later. That automatic loop is the part developers keep pointing to. Hermes also has a layered memory system. Some reports describe it as three layers. Session memory for the current work. Episodic memory through SQLite or SQLite FTS. five for past sessions and procedural memory through skills. In plain terms, it can remember conversations and tasks over time, search through past work, and build repeatable patterns for future execution. And importantly, this is done without needing some heavy cloudonly vector database setup. It is meant to run on your own machine, server, VPS, cloud environment, or even more serious infrastructure. That local first angle is another reason it is spreading. A lot of developers like the idea of an agent that lives on their own infrastructure. Your data stays under your control. The project is MIT licensed. There is no forced cloud lockin. There is no automatic dependency on one model company. Hermes can work with Open Router, Anthropic, OpenAI, Newsportal, Kimmy, Miniax, GLM, Nvidia Nim, AWS, Bedrock, Olama, local models, and custom endpoints. That model agnostic design matters because agentic workflows can get expensive very quickly. Every plan, tool, call, retry, file edit, browser step, and validation pass adds tokens. So, if an open stack can route routine, work through cheaper models, and only use expensive models where needed, that becomes a real advantage. It doesn't have to beat every premium agent at every possible task. It only has to be good enough for repeated workflows where control, cost, memory, and flexibility matter. Now, just 3 days before that ranking flip, Noose Research shipped the update that likely set the whole thing off. On May 7th, NR shipped Hermes agent version 0.13 called Tenacity. This was not a small patch. Reports describe 864 commits, 588 merged pull requests, and contributions from 295 developers in one week. That is an enormous release cadence for any open- source project, especially one sitting in a category as sensitive as autonomous agents. Tenacity added a durable multi-agent conbon system. That means Hermes can manage multiple agents or workers across a task board with heartbeat monitoring, retry budgets, zombie worker reclaim, and hallucination recovery. In normal language, it is trying to keep long-running agent work organized, prevent dead workers from silently breaking a workflow, and make the system more durable when things go wrong. That matters because long-running agents fail in messy ways. They don't just give a bad answer. They lose the thread, repeat themselves, get stuck in loops, modify the wrong file, forget the goal, call tools in strange orders, or keep working after the original objective has drifted. A durable taskboard with monitoring and recovery is the kind of feature you need when agents move from demos into actual infrastructure. Tenacity also added the /goal command, which keeps the agent locked on a long-term objective across turns. That may sound small, yet it solves a very real problem. Agents often get distracted by intermediate steps. They solve the next visible issue rather than staying aligned with the actual mission. A persistent goal gives Hermes a stronger anchor, especially for tasks that unfold across many tool calls, files, platforms, or sessions. The release also added Google Chat as its 20th supported messaging platform, showing that Hermes is expanding beyond the terminal and moving toward the same always available agent territory that made OpenClaw popular. Bigger picture is that Hermes is moving towards something closer to an agentic operating system. One article described Hermes AIOS as combining Hermes agent with the Ion UI, creating an open-source platform that can manage tasks directly on the computer. In simple terms, Hermes is moving from agent you talk to toward agent that can sit inside your computer and actually manage work. It can organize files, run code, automate dev tasks, and show its progress through a visible interface instead of acting like a mysterious black box. And most people do not want an invisible agent doing mysterious things in the background. The more powerful the agent becomes, the more you need visibility and control. A good interface becomes part of the safety model. Users need to see the tasks, the files, the actions, the failures, and the decisions. Hermes seems to be moving in that direction with the conbon system, ion UI, checkpoints, and gateway autores. And then there is migration. Hermes is not only competing with OpenClaw from the outside. It is making it easy for OpenClaw users to try the switch. During setup, Hermes can reportedly detect an existing home/openclaw directory and offer to import settings, memories, skills, and API keys. There is also a Hermes claw migrate command with dryr run previews, selective migration presets, and conflict controls. That is a very aggressive move because it reduces the friction for exactly the audience Hermes wants. People already running personal agents who may be frustrated, curious, or ready to experiment. And OpenClaw has had a rough stretch. OpenClaw has reportedly faced serious security problems including high severity CVEes, malicious entries in Claw Hub's skill repository and publicly exposed instances. That kind of turbulence matters because agents are not normal apps. They can touch files, call tools, hold keys, connect to messaging platforms, and run long workflows across real infrastructure. That does not automatically make Hermes safe. Hermes is younger, moving fast, and has already had its own issues, too. Tenacity itself reportedly closed several major security problems, including fixes around redaction defaults, roll allow lists, stranger rejection, and Oll related flows. So, the fair version is simple. Both projects are powerful. Both carry risk, and anyone running autonomous agents on real infrastructure needs to treat security seriously. There are also reports claiming around 30% of OpenClaw users have switched based on Reddit sentiment surveys, mainly citing easier setup, better memory defaults, and the self-improving learning loop. That number should be treated carefully because sentiment surveys are not the same as hard migration data. Still, it matches the broader vibe. Hermes is becoming the agent that technical users are testing when they want something that feels more adaptive. The setup also seems designed for that. Some reports say installation uses a oneline curl installer that handles dependencies like Python 3.11, Node.js, RIP GEP, and FFmpeg. After that, Hermes setup runs a wizard and can detect openclaw configuration. That matters because open- source agent tools can be powerful, yet the installation experience often kills adoption before the user even reaches the first real workflow. The broader market signal is even more interesting. Hermes topping open router matters because it shows where developer attention is moving. Even if token volume alone doesn't prove the agent is production ready, openclaw still has the bigger ecosystem, more stars, and more total usage. But Hermes is gaining momentum because it focuses on memory, reusable skills, and an agent that actually improves the longer you use it. Hermes also fits a bigger trend in AI. Agents are moving from one-off replies towards systems that remember, adapt, and act across real workflows. And the craziest part is that this still feels early. Hermes has already moved from launch to major market signal in about 3 months. It has already shipped multiple major releases, expanded platform support, introduced an automatic learning loop, added multi- aent task management, improved security, created migration paths from Open Claw, and pushed into the idea of an open-source AIOS with local control. The real test now is whether this momentum holds. If Hermes keeps this kind of usage through June, it starts looking like a real shift in developer behavior. If it fades, it still proves how quickly public rankings can reshape attention in the agent market. Either way, OpenClaw now has real pressure from a project that is not trying to beat it feature for feature, but trying to change what people expect from an agent in the first place. And that may be the most important part of this whole story. Hermes is making the case that the next major agent is not the one with the most integrations, the loudest hype, or the biggest skill store. It may be the one that watches how work gets done, turns that into memory, turns memory into reusable skills, and slowly becomes harder to replace every time you use it. That is a much bigger idea than another leaderboard win. Also, if you want more content around science, space, and advanced tech, we've launched a separate channel for that. Links in the description. Go check it out. Anyway, that's it for this one. Let me know what you think about Hermes Open Claw and whether self-improving agents are becoming the real next step. Thanks for watching and I'll catch you in the next one.