Milestone in Open Source Agents

In May 2026, the open-source agent field witnessed a milestone event. According to the latest data from the OpenRouter platform, Hermes Agent topped the global application token consumption chart with an average daily usage of 271 billion tokens, officially surpassing the long-time leader, OpenClaw. This shift in ranking immediately sparked discussions in the industry about the “Agent Era.” Many in the Chinese tech community view Hermes as a representative of the “next generation of agents” or even a “disruptive innovation.” However, a deeper technical examination reveals that Hermes is not a revolution but a systematic optimization and refinement of the “executable agent” paradigm pioneered by OpenClaw.

OpenClaw: Pioneering the Executable Agent Paradigm

To understand the innovation of Hermes, one must first grasp the significance of OpenClaw’s paradigm. The author previously analyzed this in the article “Also Discussing ‘Raising Lobsters’: What Are People Really Cheering For?” OpenClaw’s contribution lies not in a specific technical module but in its advancement of AI from a “content generation” system to a “task execution” system. The rise of OpenClaw marked a shift in the core driving force of technological evolution from “how to make AI answer questions better” to “how to make AI truly help people accomplish tasks”—a transition from “cognitive intelligence” to “executive intelligence.” Scholars have defined OpenClaw as the “scaffolding” for the practical application of AI capabilities, building a robust, flexible, and low-threshold framework on top of foundational models, allowing ordinary users to leverage top-tier models to complete complex tasks through simple natural language interactions. This modular design breaks down a grand vision into deployable and executable components: the message gateway addresses user entry issues, the Skill system standardizes tasks, and local deployment resolves data privacy concerns. This systematic construction aimed at execution scenarios forms the technical foundation of OpenClaw as a “digital service”—it does not merely add a tool-calling capability to AI but builds a scalable and iterative human-machine collaboration system from a service architecture perspective.

Hermes’ Triple Optimization: Deepening Value from “Tool” to “Partner”

Hermes optimizes the “executable agent” paradigm established by OpenClaw across three technical dimensions, elevating the agent’s value from a short-term execution tool to a long-term cognitive partner:

1. Structural Reformation of the Memory System: Hermes’ memory management is not a simple append-only record like OpenClaw’s; it is a system engineering approach that balances both “memory quantity” and “memory quality.” The design of its Memory subsystem is very restrained—two plain text files, MEMORY.md (Agent’s personal notes) with a limit of 2200 characters, and USER.md (Agent’s understanding of the user) with a limit of 1375 characters, forcing the agent to compress information and prioritize choices due to character limits. In contrast, OpenClaw’s MEMORY.md adopts a pure append mode, which can swell to tens of thousands of lines after months of operation, making information retrieval and maintenance extremely inefficient. More critically, Hermes’ overflow handling mechanism does not simply discard or compress data but involves the model actively participating in the decision-making process of what to retain or delete, internalizing “information organization” as a form of “self-reflection” for the agent. Another clever design is the session snapshot freezing mechanism: each time a session starts, the memory is loaded, and a snapshot is immediately captured, using this snapshot for the entire session instead of real-time updated active entries. This approach allows for shared prefix caching on the front end, effectively saving token costs; new content only modifies the disk and is refreshed in the next session. In essence, Hermes transitions the agent’s approach to “what to remember” from “large and comprehensive” to “small and precise” with a focus on timely forgetting.

2. Self-Evolution of the Skill System: Hermes’ core differentiation lies in its ability for agents to “self-evolve through work experience.” In OpenClaw’s environment, skills are static handwritten configuration files, meaning that an agent making a hundred deployments will repeat the same mistakes on the hundredth time as on the first. Hermes’ “learning loop” consists of a closed loop involving three key subsystems: Memory (remembering people), Skill (remembering tasks), and Nudge Engine (reminding learning). From the source code perspective, Hermes follows a closed-loop mechanism of “doing work, reflecting, refining, and reusing.” After completing a complex task—such as calling tools more than five times, encountering and fixing issues, or being corrected by users—the Review Agent automatically triggers a review process in the background, independent of the main dialogue and running as a forked subprocess, with the user completely unaware. This process limits the maximum number of iterations (default 8) and disables its own Nudge mechanism to avoid resource consumption. The review agent assesses whether it is worthwhile to create a skill based on two sets of prompts, focusing on non-trivial problem-solving processes. Once confirmed, it automatically generates a structured skill file (including name, description, applicable conditions, steps, and issue records) and writes it to the skill library. More importantly, Hermes supports progressive self-repair of skills: when an agent executes according to an existing skill and discovers missing steps or new issues, it can make precise local updates post-task completion using the _patch_skill function—utilizing fuzzy_find_and_replace to tolerate format differences and incorporating a safety scanning mechanism to ensure modifications meet security standards, rolling back if they do not. This transcends simple “creating new skills” and lays the engineering foundation for long-term maintainability of the skill system. For skill loading, Hermes adopts a progressive loading mode: the default context is extremely lightweight, containing only a lightweight index of skill names and brief descriptions; only when the agent determines that a skill is relevant to the current task does it load the complete content via skill_view. In contrast, OpenClaw employs a “heavy backpack” model, stuffing all configurations into the context at once, leading to significant token waste and diluted model attention.

3. Systematic Upgrade of Security Architecture: Security has been one of the most criticized aspects of OpenClaw. In terms of security, Hermes’ zero CVE record starkly contrasts with OpenClaw’s severe RCE vulnerabilities. In early 2026, a serious RCE vulnerability was discovered in OpenClaw, allowing attackers to hijack the localhost gateway through WebSocket for unauthorized remote control; in February, Oasis Security further disclosed that any malicious website could silently connect to the local OpenClaw gateway during developer access and brute-force passwords. Hermes, however, has built a zero-CVE security defense through a series of designs, including zero telemetry by default, automatic desensitization of confidential information, and WhatsApp’s default rejection of messages from strangers. However, it is also important to note that Hermes’ relatively lightweight user base and later release date, along with OpenClaw’s larger scale and complexity as an execution platform, objectively increase its attack surface and compliance risk exposure; in contrast, Hermes was designed from the outset with more modern security principles, avoiding some of the pitfalls accumulated during OpenClaw’s early evolution. From this perspective, Hermes’ security advantage is both a result of design optimization and a natural benefit of being a later entrant.

Examining Hermes: An Enhancer of Digital Service Innovation Rather Than a Revolutionary

Generally, we categorize service innovations into incremental, intermediate, and radical innovations, primarily judging them based on the scope of change and dual impact on customers and manufacturers. By this standard, we can clarify Hermes’ innovative position: it represents an incremental deep optimization of OpenClaw’s “executable agent” paradigm rather than a fundamental paradigm disruption.

Hermes’ evolution focuses on continuous improvement at the technical level rather than disruption at the service process, service content essence, and business model levels: it does not redefine the basic paradigm of AI-user interaction—users still issue commands to the agent through natural language, and the agent still executes tasks through tool calls; the four-layer logical architecture of the “executable agent” constructed by OpenClaw (interaction layer, cognitive layer, execution layer, memory layer) forms the foundational design and operational logic of Hermes. Hermes’ contribution lies in pushing the system’s self-evolution capability, memory efficiency, and security reliability to a new level based on OpenClaw: it upgrades the agent from a “tool that can only work” to a “partner that gets better with use”; it achieves more efficient management of tokens and memory through progressive skill loading and capacity-limited memory; and it builds a more robust data protection system based on zero telemetry and automatic desensitization designs, addressing the security pain points exposed by OpenClaw. However, Hermes has not created a brand-new “executable agent” paradigm—all its innovations essentially confirm the architecture established by OpenClaw, with innovations occurring in localized modules.

In examining Hermes, we find that it has achieved a systematic and multi-level technical enhancement on the framework of OpenClaw’s “executable agent”—it has optimized the long-term value realization of agent-human collaboration but has not altered the fundamental logic of how agents collaborate with humans. Therefore, calling Hermes the “new benchmark in the agent track” is accurate; labeling it as a “revolutionary innovation” would be an exaggeration.

Future Trends: Expanding the Sustainable Agent Ecosystem

The competition between Hermes and OpenClaw cannot be simply summarized as “which is more advanced.” From a macro ecological perspective, both represent a new stage in the agent track characterized by clear division of labor and fusion-driven development. OpenClaw’s model defines the industrial standard for “action-oriented agents”; Hermes explores the technical path of “reflective agents” based on the ecosystem inspired by OpenClaw.

From a broader digital economy perspective, the self-evolution potential of agents positions them as “digital labor” within the digital economy infrastructure. When an agent can continuously reduce users’ long-term operational costs through skill accumulation and knowledge iteration, it signifies that AI’s empowerment efficiency will grow exponentially. OpenClaw opened the first door to this trend, while Hermes has ventured further along this path.

In reality, technological development has never had shortcuts to success. Hermes stands on the “shoulders” of OpenClaw, bringing “executable AI” to new heights through more forward-looking architectural design—it is not a disruption of a technical model but a valuable ecological leap. For participants in the AI agent track, it is crucial to understand the complementary advantages of both paths and translate them into core advantages for product implementation and industry applications. After all, in the marathon of executable agent development, the true finish line is not about whose architecture is louder but about who can deliver visible value to users at lower costs and higher efficiency, building the most suitable application ecosystem.