Hermes Is Becoming the Default AI Agent

For two years the knock on AI agents has been that the ground never stops moving. The tool everyone recommends in January is abandoned by April. That churn is exactly why the latest shift is worth a look: Hermes, the agent from Nous Research, has taken the top daily-usage spot on OpenRouter, and the reasons people give for switching are unusually consistent.

What actually happened

OpenRouter ranks agents by how much real traffic runs through them, and on that board Hermes moved into first place for daily usage, running about 224 billion tokens a day to OpenClaw's 186 billion. The lead is real but narrow, and it comes with an asterisk: OpenClaw still sits ahead on all-time usage and is installed far more widely, with more than five million downloads a week. So this is a momentum shift, not a knockout. By Nous's framing, Hermes is "the agent that grows with you," and surveys of users suggest close to a third of OpenClaw's base has tried switching.

Why people are switching

The complaints about OpenClaw rhyme. Its updates ship fast, and that speed has a cost: users report that roughly a quarter of updates break something, a cron job, a webhook, a plugin, the agent's delivery of its own responses. People also describe its memory bloating over time and the agent forgetting things it was taught.

Hermes leans the other way. It ships on a more careful cycle, so there are fewer chances to break a working setup. Its memory has sensible limits so it does not balloon, and it has a self-improving loop, where the agent writes and reuses its own skills, even drafting a troubleshooting skill when something fails, instead of relearning the same task every time. The recurring praise is mundane and exactly what you want from infrastructure: it is easier to set up, and it tends to just keep working.

The honest caveats

None of this makes OpenClaw dead. It still leads on installs and lifetime usage, it moves faster, and it is more flexible through plugins and gateways. Some of Hermes's stability is simply a function of being newer and shipping less. Power users have a real gripe too: Hermes can overwrite manual edits to its skills, and plenty of people just run both, letting OpenClaw orchestrate while Hermes executes. A usage leaderboard is a popularity contest, not a proof of quality.

Here is the practical read. If you have been putting off setting up an AI agent because the tools seem to change every five minutes, this is the moment worth stopping on.

Hermes is starting to look like the default agent harness, the standard base layer people build their workflows on.

So if you are still wrestling with OpenClaw, the switch is probably worth it. And if you have never set an agent up at all, this is a reasonable, low-regret place to start.

The whole reason people delay this stuff is fear of backing the wrong horse. For once, that fear is smaller than usual.

Sources

• Hermes Agent (Nous Research, GitHub)
• OpenRouter app rankings
• Why Nous Research's self-improving agent now leads OpenRouter's rankings (MarkTechPost)
• OpenClaw vs Hermes: what Reddit says (Kilo)
• OpenClaw vs Hermes popularity and stats (Glukhov)
• The Information: Hermes eclipses OpenClaw in new GitHub contributors