Hermes Agent Tops OpenRouter Rankings: AI Agents Enter Era of Multi-Strategy Competition

【Ebrun Original】May 2026 witnessed a landmark event in the field of AI agents. According to the latest data from the OpenRouter platform, the open-source agent platform Hermes Agent has topped the global application token consumption chart with a daily average of 291 billion tokens processed. Its weekly token volume exceeded 1.75 trillion, and its cumulative token consumption surpassed 6.37 trillion, marking the first time it has overtaken OpenClaw (commonly known as "Lobster"), which had consistently held the top spot.

This is not merely a simple ranking change. Among the top five models by usage volume on OpenRouter this month, domestic models occupied the majority of positions in the core supporting tier. If the ranking reflects the trends in the AI field, then the shift in the sources of the most-called models reveals deeper changes in the industry landscape.

Hermes Agent (nicknamed "Hermes") was developed by the AI research lab Nous Research, and its rise to prominence is no accident. It has pursued a different technical path compared to OpenClaw. If OpenClaw's design philosophy leans towards connectivity to achieve "maximum coverage," then the core concept of Hermes Agent emphasizes "making your AI smarter with use." OpenClaw's core architecture is a central WebSocket gateway supporting over 50 messaging channels like WhatsApp, Telegram, Discord, and Slack, enabling agents to run on as many platforms as possible. This represents a classic "channel-first" approach—expand reach first, then optimize the experience. In contrast, Hermes Agent focuses on self-evolving persistent memory, autonomous learning, and skill self-optimization, making it appear more like an agent capable of continuous growth from execution experience.

This fundamental difference has shaped the divergent trajectories of the two projects. OpenClaw's skill files are static, requiring manual writing and maintenance by users. Hermes, however, addresses a long-standing pain point in the developer community: why can't my AI learn from experience?

Furthermore, in early 2026, OpenClaw faced a severe security crisis. In comparison, there have been no prominent reports of documented CVE vulnerabilities for Hermes Agent to date. Features like zero telemetry design and automatic sensitive information anonymization are particularly crucial for users who prioritize data sovereignty and privacy compliance.

The competition between Hermes and OpenClaw is just the tip of the iceberg; the global AI agent landscape is rapidly expanding. At the enterprise-level agent operating system tier, in the first quarter of 2026, Microsoft released Agent 365 and Copilot Cowork, OpenAI released Frontier and entered a strategic partnership with Amazon, and Google launched Gemini Enterprise while continuing to advance the A2A open protocol. The three giants are essentially focusing on the same goal: building enterprise-grade operational infrastructure for agents, addressing a series of engineering challenges such as context fragmentation, permission management, and reliable execution.

The transformation is even more profound in the search agent domain. Traditional search products are being redefined by the wave of "AI search engines" and "super search agents." General-purpose agents like Hermes can inherently perform search tasks, while specialized search layers for agents are emerging as a new infrastructure battleground. Overseas startup Tavily secured $25 million in funding, significantly boosting its valuation; its platform specializes in providing real-time web data retrieval for AI agents. Another startup, Exa.ai, built the world's first search engine specifically designed for AI agents rather than human users, offering semantic understanding-based intelligent retrieval through its own models and indexes.

Looking at the current industry landscape, AI agents are developing along two clear main trajectories. The first trajectory is "Agents that connect and achieve continuous optimization," represented by OpenClaw and Hermes. The advantage of such agents lies in their ability to become more capable with higher usage frequency. However, the challenge is the extremely high demand for instruction-following accuracy in the underlying models, million-level context processing capability, and reasoning stability. The second trajectory is the "Integration of AI Search Engines and Agent Search Layers." Traditional search engines optimize for keyword matching and ad clicks, whereas AI agents require semantic understanding, real-time data, and reliable sources. This integration signifies a new stage for search technology, moving from information matching towards intelligent problem-solving.

In fact, these two trajectories are not mutually exclusive; they are converging. An ideal AI agent should both "get smarter with use" (self-evolution capability) and "find and retrieve information accurately" (intelligent search capability), and also "execute successfully" (automated execution capability) after planning complex tasks. This is the core vision of what the industry calls "Agentic AI"—the evolution of AI from a conversational tool into a task solver possessing both execution and judgment capabilities.

Previously, the competitive focus in the AI field was on model parameters, benchmark scores, and launch hype. Now, an increasing number of products are being tested by real-world usage volume. Token consumption does not equate to success, but it indicates at least two things: first, there are real users making frequent calls; second, the application is handling complex tasks, not just superficial Q&A. Overall, the development paths and architectures of AI agents are revealing mature contours. As the global AI agent architecture rapidly takes shape, the AI industry is undergoing swift and profound transformation.

We are currently compiling the industrial ecosystem maps for OpenClaw and Hermes. If you have noteworthy products or achievements, please contact us at:

huangbin@ebrun.com