Issue 354

According to The Batch, the open-source agent Hermes Agent, launched by Nous Research in February 2026, recently surpassed the previously dominant OpenClaw in daily token consumption on OpenRouter. This metric reflects actual user usage, but some users note that Hermes Agent has lower token efficiency, meaning it may consume more tokens to complete the same tasks.

Hermes Agent's functionality largely overlaps with OpenClaw, with core differences in memory architecture and the ability to autonomously build skills. It is designed to run locally or in the cloud, supports multiple large language models, and integrates with about 20 messaging services. By using a local model or browser login to generate new access tokens, users can run it without storing API keys.

Hermes Agent uses standard SKILL.md format instruction files, comes with built-in skills, and can fetch additional skills from the Skills Hub (currently much smaller than OpenClaw's community skill library). However, its unique feature is the ability to automatically create new skills: when the agent works on a problem for an extended period or fixes an error and determines the task is successfully completed, it calls a tool to generate a skill file.

To prevent uncontrolled skill growth, the background system Curator archives skills unused for over 90 days and uses an LLM to decide whether each skill should be kept, merged, or archived. For memory, Hermes Agent maintains two general memory files (user preferences, workflows and lessons learned), adds content via built-in memory tools, and checks whether content is worth adding and which file it belongs to before adding. When a file exceeds length limits, it merges related entries. Additionally, it maintains a searchable conversation database and supports external memory provider Honcho for identity analysis.

Hermes Agent's agent loop is similar to OpenClaw's: it assembles a prompt based on personality, instructions, tools, skills, memory, user knowledge, and conversation history; if the prompt exceeds limits, it asks the LLM to summarize old messages; after sending the prompt, it calls tools, skills, or directly responds to the user. The loop repeats until a user response is generated.

Persistent goal tracking allows users to specify goals in messages; after responding, the agent calls a judge model to evaluate whether the goal is achieved. If not, it continues working until the goal is met or the maximum number of rounds is reached. Anthropic Claude Code, OpenAI Codex, and OpenClaw (via plugins) also offer similar functionality.

Although Hermes Agent leads in token consumption, its lower token efficiency may affect actual user experience. The Skills Hub is much smaller than OpenClaw's community skill library, meaning fewer pre-built skills are available. While automatic skill generation is innovative, the effectiveness of the Curator system (e.g., the accuracy of LLM judgments) has not been independently verified.

Persistent goal tracking relies on a judge model, whose reliability depends on model quality. Moreover, performance comparisons between Hermes Agent and OpenClaw on specific tasks (e.g., success rate, latency) have not been publicly benchmarked.