Hermes vs. OpenClaw, Cybersecurity Alarms Ring, More-Interactive Conversations, Can Agents Do Human Work?

There's a new leader in the AI-agent popularity contest, and it snuck up fast. Hermes Agent, an open-source project from New York's Nous Research launched only in February, has now pulled ahead of OpenClaw on OpenRouter's leaderboard for daily token consumption. That's a real signal of adoption, not just hype, even though some users grumble that Hermes burns through tokens less efficiently than its rival.

What sets Hermes apart isn't a flashier chat interface or a bigger model roster, though it does plug into roughly 20 messaging apps and run locally or in the cloud. It's the memory and skill system underneath. The agent writes standard SKILL.md instruction files on its own whenever it finishes a hard task or fixes a bug, essentially banking know-how for next time. A background process called Curator then cleans house, archiving anything unused for 90 days and asking an LLM to decide whether skills should be merged, kept, or retired. That's a small but meaningful design choice: most agent projects let their skill libraries balloon into clutter, and Hermes builds in janitorial work from day one.

Memory works similarly. Two running files track user preferences and lessons learned, and the agent checks before adding anything, so it won't jot down a vague or duplicate note. It can also lean on an external service, Honcho, which studies conversation patterns to infer what a user actually wants. Add persistent goal tracking, where a judge model checks whether a stated objective got done and keeps the agent grinding until it does or hits a turn limit, and you get something closer to a colleague who remembers your quirks than a chatbot that resets every session.dilluter This isn't happening in a vacuum. Claude Code and Codex made coding agents mainstream through 2025, and OpenClaw turned the idea of an always-on personal agent into a genuine open-source phenomenon earlier this year, security warts and all. Its creator later joined OpenAI, which tells you something about how seriously the big labs are taking this category. Hermes is part of a wave of OpenClaw-inspired projects, but it's the self-improvement angle, agents that get measurably better at your specific workflows over time, that looks like the next battleground.

The bigger point is that general-purpose agents are converging on a shared feature set, memory, skills, tool use, while still leaving plenty of room for someone to out-engineer the pack on any single piece. Hermes doesn't need to be flashy. It just needs to remember what worked last Tuesday.