·Preface:

Throughout the history of technological innovation, a recurring pattern emerges: when multiple products with similar functions but divergent design philosophies appear in the same domain, it often signals that the field is undergoing rapid evolution.

Developer Migration: Driven by Precise Pain Point Resolution

This open-source project, affectionately dubbed '[Hermès]' by Chinese developers, underwent nine months of meticulous polishing. Within less than two months of its launch, it amassed 66,000 stars on GitHub, with over 8,800 forks.

What excited Chinese developers even more was its seamless native integration with WeChat. On its launch day, the official Chinese announcement post garnered 549,100 views.

Overnight, a significant number of developers abandoned OpenClaw in favor of Hermes, drastically reshaping the AI Agent landscape in just two months.

OpenClaw initially gained popularity with its minimalist installation process and multi-channel access capabilities. However, as its user base expanded, underlying issues began to surface.

Some developers observed in comparative tests that OpenClaw was prone to frequent crashes, making daily use cumbersome. The upgrade process was also error-prone, with each update causing anxiety among users.

The industry also criticized its inefficient token consumption. OpenClaw’s context management was wasteful, often breaking down user queries into multiple low-value tool calls. Consequently, the actual reasoning cost for a single task could skyrocket to dozens of times the subscription price.

Its Skill system also posed significant barriers. Users had to manually create files, install authorizations, and restart the Gateway process for new skills to take effect, resulting in high maintenance costs.

Hermes precisely addressed these core pain points. Under the same task and LLM model, Hermes’ token consumption was merely a quarter of OpenClaw’s, significantly reducing user costs.

With an entry threshold of $3.99, users could instantly launch Hermes on Agent37, integrating over 1,000 applications with pre-configured LLM models, truly achieving plug-and-play functionality.

Native WeChat integration further propelled Hermes’ popularity in the Chinese market. Hermes leveraged Tencent’s official iLink Bot API for integration, avoiding the need for reverse engineering or unofficial clients.

It also supported mainstream domestic IM tools like DingTalk, Feishu, and WeCom, catering to the most critical usage scenarios for Chinese users.

Beyond Functional Overlap: Fundamentally Different Architectures

While many perceive OpenClaw and Hermes as direct competitors, their system architectures are fundamentally different, despite some functional overlaps.

OpenClaw’s core is the Gateway control plane, resembling an Agent-version personal communication and device control hub.

Hermes’ core, in contrast, is a learning-based Agent Runtime, treating the Agent’s execution process as a long-term asset.

① Skill System: OpenClaw boasts a comprehensive skill system with over 50 built-in skill directories, supporting hierarchical loading and governance. However, every step of the skill lifecycle requires user intervention.

Manual file creation, authorization installation, and process restarts are necessary for skills to take effect. Even skill loading, regardless of task relevance, dumps everything into the context, leading to token waste.

Hermes transforms the skill lifecycle into an automatically evolving closed loop.

During daily task execution, if an Agent calls a tool more than five times, completes error repairs, or receives user corrections, the system triggers hard rules.

It silently packages the successful workflow into a local SKILL file, often without the user even noticing.

For similar tasks in the future, Hermes matches skills through four layers of progressive loading: first by name description, then expanding the full content as needed, ensuring both matching efficiency and token control.

More critically, Hermes features offline evolution capabilities. It incorporates an offline batch evolution system based on the GEPA algorithm, using reflective mutation, Pareto frontier selection, and natural language feedback as its three core pillars to regularly optimize existing skills.

Unlike the industry’s mainstream reinforcement learning approach, this system achieves skill iteration and optimization without gradient updates, relying solely on the model’s reflective capabilities and evolutionary algorithms for higher sample utilization efficiency.

② Memory System: OpenClaw adopts a ‘[File as Memory]’ approach, storing core memories in files like SOUL.md, USER.md, and MEMORY.md.

It only triggers hidden rounds to record conversation highlights into files when the context is nearly full and about to be compressed.

Hermes’ memory system is a three-tiered structured framework, progressing from temporary session memory to cross-session persistent memory and finally to reusable skill memory.

Its writing mechanism is entirely proactive. The system sets a hard nudge mechanism, triggering a reflection instruction every 15 conversation rounds to force the Agent to review the dialogue, extract user preferences, and write them into persistent memory.

It also natively incorporates SQLite FTS5 full-text search capabilities, enabling rapid retrieval of full historical sessions without additional vector service configuration, significantly improving memory recall efficiency and accuracy.

It supports third-party memory backends like Honcho and MEM0, refining users’ fragmented expressions into structured insights through asynchronous dialectical reasoning.

③ Harness Design: Using Deterministic Rules to Counter Model Uncertainty

OpenClaw’s Harness design offers users extensive configuration freedom, delegating many decisions and judgments to users and large models.

Hermes’ Harness design, however, centers on using deterministic code rules to counter the uncertainty of large models.

Tool call counts, conversation rounds, and error repair scenarios—all conditions triggering skill generation and memory writing—are hardcoded. Once conditions are met, execution occurs immediately without fuzzy judgments by the large model.

Core processes like context compression, security approvals, and plugin management are also implemented with fixed rules, sacrificing some flexibility for system stability and determinism.

For example, in context management, when the conversation reaches 85% of its threshold, Hermes’ ContextCompressor directly replaces old tool outputs with placeholders using fixed string replacement logic—crude but absolutely safe.

At the memory level, it uses frozen snapshots, loading memories into system prompts at startup without refreshing them midway, only updating upon the next restart. This sacrifices real-time performance but ensures stable prefix cache hit rates, directly reducing token input costs by about 75%.

Why Both OpenClaw and Hermes Coexist

One is OpenClaw, with over 350,000 GitHub stars, affectionately called '[Little Lobster]' by the Chinese community.

The other is Hermes Agent, which amassed 60,000 stars within two months of its launch, dubbed the '[Hermès of the Lobster World]'.

OpenClaw follows a configuration-driven approach. Users set its personality, customize rules, and install Skills, and it operates according to these configurations.

Its strength lies in controllability, predictability, and suitability for enterprise deployment. To enhance its capabilities, users configure, optimize, and maintain the Skill ecosystem themselves.

This resembles traditional IT systems—the system is static, but humans are dynamic. Users inject whatever configurations they need for desired functionality.

OpenClaw represents the classic community-driven model in the open-source world, initiated by individual developers and rapidly iterating through the collective wisdom of the global community, fostering a thriving ecosystem, much like Linux in its heyday.

Hermes, on the other hand, follows a self-learning-driven approach. It operates according to configurations but also automatically summarizes experiences after execution, precipitating them into memory and Skills for use in future tasks.

After completing a complex task, it automatically reviews the execution process: which operations succeeded, which failed, which took too long—then generates a structured skill file stored locally.

Hermes Agent represents the professional laboratory-led open-source model, where teams with deep technical expertise create core architectural innovations and amplify their value through open-source ecosystems, much like MySQL in its early days.

In short, OpenClaw requires users to '[raise]' it, while Hermes grows on its own.

OpenClaw and Hermes address problems at different levels. OpenClaw is more like an executive digital employee—users tell it what to do, and it executes tasks stably and efficiently.

Hermes is more like an evolving digital partner—users teach it once, and it learns for a lifetime.

These two positioning cater to different usage scenarios. For example, in a customer service scenario, enterprises may prioritize compliance and traceability over the AI’s intelligence.

Here, OpenClaw’s configuration-driven advantage shines—all Skills are manually reviewed, every operation is recorded, and issues can be traced to specific configuration changes.

But for individual users seeking a 24/7 personal assistant, Hermes’ advantages are more apparent.

Just like Chrome vs. Firefox in the browser market or Windows vs. macOS in the operating system market, different products cater to different audiences, and different philosophies serve different needs.

Conclusion:

A millennium ago, the ‘[Rivalry Between Zhou Yu and Zhuge Liang]’ left behind the timeless lament of ‘[Since Zhou Yu exists, why must Zhuge Liang also be here?]’. Yet, it was this very confrontation that made both men the brightest stars of the Three Kingdoms era.

The same is true for the ‘[Rivalry Between Zhou Yu and Zhuge Liang]’ between OpenClaw and Hermes Agent. They serve as mirrors to each other, exposing each other’s shortcomings and pushing each other to become better.

Partial References: Architect: 'OpenClaw vs Hermes: A Deep Dive into Two Universal Agents', Mr. K: 'Hermes Goes Viral! Completely Overtakes OpenClaw', New Intelligence: 'Lobster Steps Aside! Silicon Valley’s Top AI '[Hermès]' Breaks into WeChat Overnight, Tops Global Charts', Datawhale: 'Latest! A 10,000-Word Overview of the Harness Revolution!', Boyang: 'Why Did Hermes Replace OpenClaw in Just Two Months?'