When OpenClaw launched in early 2026, it spread fast. The open-source personal AI agent accumulated more than 200,000 GitHub stars, built a thriving skills marketplace, and made the case that always-on AI agents running through messaging platforms like WhatsApp and Telegram were ready for everyday use. Its inventor eventually joined OpenAI.
Now it has serious competition.
Hermes Agent, an open-source agent released in February by New York-based AI lab Nous Research, recently surpassed OpenClaw on a key usage metric: daily token consumption, as tracked by the AI model platform OpenRouter. For an agent ecosystem that was barely a year old, that’s a meaningful shift in the benchmark.
But the more interesting story isn’t the leaderboard. It’s what Hermes Agent does differently — and what that signals about where general-purpose agents are heading.
Self-Improving by Design
Most AI agents today are static. You configure them, give them tools, and they execute. Hermes Agent takes a different approach: it builds new skills as it works.
In the SKILL.md format used by many agentic frameworks, skills are instruction files that tell an agent how to accomplish a task using tools like bash scripts, web search, or database queries. Hermes Agent ships with built-in skills and can access additional ones from Skills Hub, a marketplace that’s smaller than OpenClaw’s crowd-sourced library but growing.
What sets Hermes Agent apart is that it also creates skills on its own. When the agent works through a complex problem or resolves an error — and judges the result successful — it automatically generates a new skill to encode that behavior for future use. A background system called Curator manages the resulting library, archiving unused skills after 90 days and periodically using an LLM to decide whether to keep, merge with similar ones, or retire skills.
The result is an agent that gets incrementally better at the specific tasks you actually use it for.
A More Sophisticated Memory Model
Hermes Agent also rethinks how agents remember context. It maintains two persistent memory files: one for user preferences and one for workflows and lessons learned. Before adding anything new, the agent checks whether a similar memory already exists, whether the new memory is specific enough to be useful, and whether it would push the file past a preset length. When it does, it automatically merges related entries rather than just appending.
Beyond those files, Hermes Agent maintains a searchable conversation database. It also supports external memory providers like Honcho, which analyzes every message to build a model of user preferences, goals, and patterns over time.
This is a meaningful departure from most current agents, who treat each session as a fresh start. Hermes Agent is designed to accumulate context across interactions — not just remember what you said yesterday, but recognize patterns in how you work.
Goal Tracking and Persistent Execution
Hermes Agent also adds a persistent goal-tracking loop. Users can define a goal in a message, and after the agent responds, a separate judge model evaluates whether the goal was actually completed. If not, the agent keeps working — up to a configured maximum number of turns. Similar capability exists in Claude Code, OpenAI Codex, and OpenClaw via a plugin.
For users who want to hand off a task and come back later, that’s a practical feature with real implications for how much autonomy they can trust these agents to handle.
The Bigger Picture
Mitch Ashley, VP and practice lead for software lifecycle engineering at The Futurum Group, sees Hermes Agent as confirmation of a broader pattern. “OpenClaw signaled the start of a category, and the variants that followed confirm personal AI agents are converging on experience-accumulating designs,” he said. “Hermes Agent advances that pattern with self-generated skills, persistent memory, and cross-session goal tracking, turning each successful execution into a reusable capability.”
That convergence has implications for enterprises that go beyond feature comparisons. “Agents that modify their own skill libraries and accumulate context create verification obligations; stateless assistants do not,” Ashley added. “Enterprises must observe and constrain that behavior before extending autonomy, because the visibility ceiling sets the deployment ceiling.”
It’s a pointed reminder that the same capabilities making these agents more useful also make governance harder. An agent that builds its own skills and retains cross-session memory is more powerful than one that doesn’t — and less predictable in ways that matter to security and compliance teams.
Hermes Agent runs locally or in the cloud, supports a wide range of large language models, and integrates with around 20 messaging services. It connects to integrated development environments via the Agent Communication Protocol. Like OpenClaw, it’s not tied to a specific model, platform, or skill format — a design choice that keeps options open as the underlying AI landscape continues to shift quickly.
The fact that Hermes Agent has outpaced OpenClaw on daily token consumption in just a few months is partly a reflection of its growing community. But it’s also a signal that users are responding to agents that do more than execute tasks. They want agents that adapt, learn, and get better over time.
That’s a different kind of tool. And it’s becoming the baseline.
