TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Behind Hermes' Rise to the Top: A Web3 Team's Journey of Advancement

Behind Hermes' Rise to the Top: A Web3 Team's Journey of Advancement

2026.08.11
Share
Hermes has at least proposed a more mature Crypto AI path: making Crypto the organization and infrastructure, rather than the product interface users must face.

Author | Jacob Zhao @ IOSG

Hermes' phenomenal growth does not stem from exclusive technology that OpenClaw cannot replicate in principle, but because during the critical window when the personal Agent category was taking shape, it most accurately closed a set of "Challenger Growth System": taking over the user pool already educated by OpenClaw, and establishing an experience difference of "Delegatability" (Delegation Trust) that is more real than the "Self-improvement" narrative. As professional execution Agents become increasingly powerful, users still need a long-term online, trustworthy Chief Steward.

Opening OpenRouter's public application leaderboard, Hermes Agent ranks first across the entire platform with 30.5 trillion Token usage, while simultaneously ranking first in four categories: Productivity, Coding Agents, Personal Agents, and CLI Agents, leading known Agents like OpenClaw and Claude Code by a significant margin.

▲ Figure 1 · Hermes Agent Historical Data Snapshot on OpenRouter (Captured on August 4, 2026; dynamic page data may change over time)

Although OpenRouter's statistical scope cannot cover the industry-wide Token consumption of direct official API connections (such as native Claude or Codex subscriptions), as the current largest-scale AI large model routing and aggregation platform globally, its leaderboard holds strong "barometer" significance. Although at the level of high-end professional tasks, a large number of users' core business workflows—complex code generation, architecture design, high-value data analysis—still flow to Claude Code and ChatGPT, Hermes maintains advantages in usage scenarios such as backend automation, message entry response, long-term online listening, and lightweight task scheduling. As an Agent product built by a Web3 team, Hermes has achieved communication, community, and usage intensity success far beyond expectations, prompting us to pay attention to:

  • Why was Hermes able to achieve a comeback in OpenRouter inference calls?
  • Where exactly lies the real dividing line between it and OpenClaw?
  • In its relationship with Claude Code and Codex, how does Hermes maintain "differentiated coexistence" rather than "head-on competition"?

From Development Framework to Personal AI System—The OpenClaw Path

Why Early Agent Frameworks Did Not Produce Consumer Products

Before OpenClaw appeared, although the Agent field already possessed mature infrastructure, there were fundamental limitations: its unit of adoption was "development project enterprise workflows," not "individual users." The common characteristic of early frameworks was being developer-oriented, outputting code or configuration—they built the infrastructure for Agents, but did not deliver the Agent itself. The excessively high engineering threshold caused them to remain stuck in the "developer tool" stage, lacking a productized closed loop to transform technology into "personal exclusive assets," leaving the "personal Agent product layer" directly facing end users almost blank.

▲ Figure 1 · Six-Layer Structure of Agent Technology Stack (Model Layer → Protocol Layer → SDK Development Framework Layer → Orchestration Runtime Layer → Execution Infrastructure Layer → Deployment Governance Layer)

▲ Figure 1 · Hermes Agent Historical Data Snapshot on OpenRouter (Captured on August 4, 2026; dynamic page data may change over time)

What OpenClaw Really Changed

OpenClaw did not reinvent Agent Loop or task scheduling technology at the bottom layer; its core contribution lies in systematic encapsulation at the product level. LangChain solves "how to build an Agent," while OpenClaw solves "how to own an Agent." It skipped the middle layers of the technology stack, integrating scattered framework capabilities into a complete product that individuals can directly configure and use long-term, achieving a fundamental shift in the unit of adoption from "development projects" to "individuals," specifically reflected in six dimensions of product innovation:

  • Identity Personification: Endowing Agents with continuous names and identity recognition, breaking the tool-like feeling of stateless API calls.
  • Entry Dailyization: Using high-frequency communication software like Telegram/WhatsApp as interaction interfaces, replacing complex command lines or IDEs.
  • State Residency: Running long-term online as a background process, achieving a leap from passive "on standby" to active "presence."
  • Permission Entity: Deeply incorporating the user's file system, browser, terminal, and real-world action capabilities into the Agent's operational boundaries.
  • Capability Extensibility: Through Skills, Memory, and community plugins, precipitating processes into reusable capabilities to expand action boundaries.
  • Mindset Ownership: The core shift—users transition from "using an AI tool" to "owning an exclusive digital partner."

Why the Lobster Craze Did Not Form a Second Mindset

The hype around OpenClaw spawned a large number of imitations. These products solved real user problems: cumbersome installation processes, difficulties in environment configuration, lack of channels like WeChat and Feishu, compatibility with domestic models, rapid deployment of cloud hosts, enterprise permission management, automatic updates, and security isolation. They all have their own user groups and reasonable business logic. But almost none formed an independent brand mindset—the reason lies in that they answered the question "how to use OpenClaw more easily," rather than "where should personal Agents evolve after OpenClaw." The position of narrative challenger is extremely scarce in the entire personal Agent market.

Why Hermes Ultimately Came from Behind

Nous's Models, Community, and Crypto-native Background

Nous Research originated from the Discord open-source AI research community in 2022 and formally completed corporate operation in 2023. The core founding team includes Jeffrey Quesnelle, Karan Malhotra, Teknium, and Shivani Mitra. Its business covers:

  • Hermes Model Series: Nous's most representative open-source model brand, long-term focused on model post-training, instruction fine-tuning, and Agent capabilities, establishing a vast developer adoption base on Hugging Face.
  • DisTrO (Distributed Training Over-the-Internet): Significantly reduces cross-node communication overhead in distributed training, making it a more feasible engineering path for cross-region, heterogeneous hardware to participate in collaborative training under internet bandwidth conditions.
  • Psyche Decentralized Training Network: Further networks DisTrO, coordinating global distributed computing nodes through Solana, enabling GPUs in different network and hardware environments to participate in large model training together.
  • Hermes Agent: A personal Agent product launched by Nous for end users, integrating Hermes models, tool calls, Memory, Skills, message channels, and long-term running capabilities into a resident Agent.

In April 2025, Nous Research completed a $50 million Series A financing round led by Paradigm, with a post-investment token valuation of $1 billion. Prior to this financing round, the company had cumulatively completed approximately $20 million in early financing, with investors including Distributed Global, North Island Ventures, and Delphi Digital, among other well-known institutions.

Nous has built a technical closed loop of "Hermes (model capabilities), DisTrO (distributed training), Psyche (decentralized computing power network), and Hermes Agent (personal terminal product)." The release of Hermes Agent was not a temporary Fork chasing hype, but a strategic-level extension from the supply side (data, models, training, open weights) to the demand side (real users, tasks, workflows) after Nous completed long-term precipitation—this provided a deeper starting point for establishing differentiation than ordinary imitation disks.

OpenClaw's Operations Pain Points Become Hermes' Growth Engine

Hermes and OpenClaw do not differ significantly in underlying encapsulation (model + tools + Memory + scheduling). Its phenomenal explosion did not rely on technological generation gaps, but precisely closed a systematic growth causal chain: directly taking over users who were already educated by OpenClaw and suffered from operations pain points through seamless migration tools, constituting the earliest core growth engine.

Product Power Leap: Establishing "Delegation Trust"

Hermes' core product hypothesis is to solve "operations responsibility transfer," promising "errors are absorbed and repaired internally by the system":

  • Reliability Trust: Ensure task continuation and failure recovery (persistent Kanban, /goal mode, tool self-healing).
  • Security Trust: Prevent unauthorized access, accidental deletion, or data leaks (Approvals workflow, sandbox, strict permission boundaries).
  • Verifiable Trust: Prove tasks are truly completed (Completion Contract and Grounded Citations).

Concept Analysis: "Self-improvement" (Narrative Advantage) vs "Autonomous Recovery" (Experience Difference)

In Hermes' product narrative, there are significant differences in product value between "Self-improvement" and "Autonomous Recovery":

  • Self-improvement: Essentially process adaptation based on Memory and Skills. Given competitors possess similar infrastructure, its differentiation lies more in being the first to integrate into a default system with lifecycle management, occupying the narrative advantage of the "growable" mindset, rather than proven, insurmountable technical barriers.
  • Autonomous Recovery: This is the experience difference most worth verifying currently. Thanks to structured error returns and Provider automatic Fallback, Hermes can digest failures within the system. This system-level stability of "not frequently disturbing users" is a more direct, perceptible product power difference.

Architecture Dividend: Delegation and Supervision Capabilities for Professional Agents

Hermes' core value does not lie in personally executing all professional tasks, but in acting as the Orchestrator layer to undertake requirement completion, task decomposition, routing monitoring, and final acceptance. By delegating external CLIs like Claude Code/Codex to execute specific work through built-in Skills, the community has precipitated a practice paradigm of "Hermes Control + External CLI as Worker" (such as /goal mechanism and oh-my-hermes collaboration tools), reflecting its architectural advantage of pushing the upper limit of task complexity by scheduling professional Agents.

From Crypto-native to Crypto-invisible: Hermes' Web3 Backend Operating System

Attributing Hermes' success simply to a "Web3 background" is an oversimplification. What Web3 provided Nous is a set of "organizational operating system" that other AI startup teams find difficult to obtain simultaneously, enabling it to切入 the mainstream market with the smooth experience of a standard AI product:

  • Patience of Venture Capital: Crypto-Native capital supports long-term, high-uncertainty, and multi-route parallel investments, enabling Nous to layout models, training, Runtime, and Cloud simultaneously without needing to converge on a single revenue verification too early.
  • Ready-made User Market: Provided a Crypto AI user group familiar with Telegram, servers, APIs, and self-hosting, significantly reducing cold-start education costs, and spawning high-intensity usage, tutorial dissemination, and Skills contributions.
  • User Sovereignty Values: Upholding self-hosting, openness, portability, and anti-platform lock-in orientations, directly implemented as MIT License, multi-Provider support, BYOK, and underlying architecture where Memory/Skills are portable.
  • Community R&D and Verticalization: Relying on global remote collaboration and open-source culture, users spontaneously become Contributors, Skill authors, and product designers for vertical scenarios.

Hermes hardly exposes Crypto to the user front end. Using its Agent, Memory, Skills, and automation capabilities does not require connecting a wallet, buying Tokens, or understanding Solana. At the same time, Paradigm capital, Psyche, distributed training, and the Crypto AI community still exist in the product backend. This forms a product morphology that can be summarized as "Crypto-native in organization, crypto-invisible in product"—retaining the most valuable parts of Crypto at the organizational level (capital, global community, user sovereignty, and coordination capabilities), while removing the parts most likely to hinder mainstream adoption at the product level (wallets, Tokens, speculative narratives, and on-chain operation friction).

Why OpenClaw Rejects Crypto, Why Hermes Hides Crypto

The superficial opposition between OpenClaw and Hermes on the Crypto issue is not an ideological battle of "rejection" versus "embrace." From the product result perspective, both reflect orientations of open source, user control, and reducing platform lock-in; the difference lies in Nous further using crypto-economic mechanisms for distributed training coordination, while OpenClaw mainly achieves user sovereignty through Local-first architecture:

  • OpenClaw (Local-first Sovereignty): Resists financial speculation, defending "Local-first" sovereignty. Due to early encounters with fake coin scams, it adopts "zero tolerance" towards Crypto. Through pure open source and local operation, it achieves non-blockchain defense of user sovereignty, firmly refusing financialization at the product level.
  • Hermes/Nous (Cryptoeconomic Sovereignty): Engineering-oriented, Crypto only as a underlying coordination tool. Introducing blockchain is a pragmatic choice to cope with engineering challenges (e.g., Psyche network uses Solana to coordinate heterogeneous computing power), rather than building financial narratives for end users.

Hermes' Advanced Mode—From Personal Agent to Task Chief Steward

This section hopes to answer a more fundamental question: When Claude Code and Codex can already complete most professional execution tasks with high quality, what is the reason for Hermes' existence as an independent product?

  • Model A: Direct Collaboration Type (Limited Gain): Users are accustomed to manually generating Prompts in LLM and handing them over for execution, manually transporting results and reviewing. Although single output quality is high, all project management and multi-Agent coordination work must be borne. For such hands-on users, Hermes' automation is instead seen as an "intermediate layer increasing opacity," failing to effectively reduce burdens.
  • Model B: Delegated Management Type (Significant Gain): Users take Hermes as a resident controller, only issuing final goals. Hermes is responsible for task decomposition, delegating sub-tasks, tracking GitHub/CI status, and automatically triggering rework. Community practices (such as oh-my-hermes) show that Hermes' core value lies precisely in replacing tedious cross-Agent coordination and project management work.

Under this framework, Hermes and Claude Code/Codex are not a substitution relationship, but a layered relationship: the latter provides execution quality at the third layer, while the former provides continuity, cross-session state, and cross-Agent coordination at the second layer. Hermes' value is not evenly distributed across all users, but may be highly concentrated in the advanced user group involving cross-Agent, cross-system, long-duration asynchronous tasks. This judgment is more precise than the general "personal Agent second mindset has formed," and is also more suitable for guiding commercialization and product priorities.

▲ Figure 2 · Hermes Agent Technical Architecture Panorama (User Entry → Gateway → Control Core → Provider Layer → Execution Layer → Orchestration Layer → State Layer → Governance Layer)

Based on official documentation and community research, the Hermes Agent technical architecture panorama framework covers the full link from user interaction to learning governance:

  • System-level Support for Autonomous Recovery: The "Control Core" explicitly includes Context compression, Provider Fallback, and interrupt state saving, providing a technical basis for failure recovery and system self-healing capabilities when tasks fail.
  • Execution Logic of "Delegation Rather than Substitution": The "Tools and Professional Execution Layer" places external CLIs like Claude Code and Codex on par with Hermes native tools (Terminal, Browser, etc.), confirming its positioning as a scheduling hub.
  • Governance Attributes of "Self-improvement": The "Learning, Maintenance, and Governance Layer" includes nodes such as Curator, Skill/Command Approval, indicating that its experience precipitation is a governance process with human intervention mechanisms, rather than a fully automatic black box.

Business Model—Who Pays for "Hermes"?

If only its Token costs are compared with direct subscriptions to Claude Code/Codex, misleading conclusions will be drawn. Because this algorithm ignores Hermes' core value: replacing user hands-on project management, context transportation, and cross-Agent coordination work.

User Value Formula Hermes User Value = Saved Manual Coordination Time + Asynchronous and Unattended Value + Cross-system Automation Benefits − Token and Tool Costs − Manual Intervention Costs − Failure and Security Risks

Therefore, Hermes' economics are not absolute, but highly depend on the user's "Delegation Depth":

  • High Delegation Depth (Economics Hold): If Hermes can transform tasks that originally required hours of manual monitoring into truly unattended execution, even if Token costs are slightly higher, its overall time cost and efficiency benefits remain positive.
  • Low Delegation Depth (Economics Collapse): If users still need to frequently intervene to correct errors and fight fires, Hermes becomes a pure Token consumer and failure amplifier.

This mechanism precisely explains why different user groups have diametrically opposite evaluations of Hermes' economics, and also hints to us: the key to verifying its business logic lies in quantifying "unattended completion rate" and "manual interventions per task," rather than simply comparing model API unit prices.

Commercialization Base: Nous Portal and Hermes Cloud

Hermes Agent adopts the MIT protocol open source, positioned as an ecosystem growth engine. The true commercialization closed loop focuses on Nous Portal, whose core value proposition is "one subscription, integrating multiple types of API keys," covering three modules:

  • Model Routing: Aggregating 252 models (providing inference through OpenRouter and direct Provider connections).
  • Tool Gateway: Built-in high-frequency tools such as Firecrawl (web search), FAL (image generation), Browser Use (cloud browser), Modal (sandbox execution), and OpenAI Audio (TTS).
  • Hosting Services: Ready-to-use Hermes Cloud instances (charging daily running fees, excluding inference and tool call fees).

Nous' actual revenue highly depends on user usage paths, currently showing obvious structural differentiation:

Open Source and Commercialization, Will Hermes Become the "Linux of the Agent World"

Hermes' MIT open-source strategy drives explosive growth while also constituting a structural constraint on commercialization. The self-hosted free model requires its paid version to possess irreplaceable extra value, but a clear differentiated monetization path has not yet formed. A deeper risk lies in "Value Interception": If Hermes continues to be widely integrated by cloud vendors as an optional Runtime, it may reenact the classic dilemma of Linux or K8s, where core commercial value is intercepted by cloud vendors providing computing power and hosting. The MIT protocol, while exchanging for ecosystem prosperity, also means giving up absolute control over distribution channels. As long as users can freely choose "self-hosting + own API" or "third-party cloud deployment," huge usage volume cannot be forcibly converted into direct revenue, facing Nous with the severe test of "niche elevation" mismatched with "actual commercial returns."

Agent Niche—Tripartite Landscape of Personal Chief Steward, Professional Tools, and Big Factory Claw

OpenClaw, Hermes, and Claude Code, Codex, and big factory hosted products differ significantly in target users and core propositions, belonging to different细分 tracks. To clarify the current market landscape, the AI Agent panoramic core competition matrix is as follows:

Hermes did not pursue the general mass market, but precisely cut into four categories of high-density Power Users, constituting the cornerstone of its phenomenal dissemination:

  • Self-hosting and Infrastructure Players: Familiar with VPS/Docker/SSH, viewing Hermes as a natural control layer for existing infrastructure.
  • Multi-model Arbitrageurs: Refusing single vendor lock-in, accustomed to dynamically scheduling frontier or local models based on tasks.
  • Multi-Agent Coordinators: Urgently need to automate orchestration of complex串联 workflows across platforms and tools.
  • Open Source and Crypto AI Community: Highly identify with user sovereignty and decentralization concepts, deeply resonating with Nous' organizational culture.

Although the base of this group is small, they possess extremely high Token consumption, code contribution, and technical evangelism capabilities, serving as the core engine driving early word-of-mouth dissemination.

Claude Code/Codex: Both Supplier and Threat

#Short-term Symbiosis: Pushing Execution Upper Limits

In actual workflows, Hermes acts as the control layer, calling Codex (code implementation) and Claude Code (architecture and review) through delegation mechanisms. The stronger the underlying professional Agents, the higher the upper limit of task complexity Hermes can deliver, forming a symbiotic relationship of "Hermes responsible for routing and acceptance, professional Agents responsible for execution."

#Long-term Potential to Swallow Hermes' Independent Value

Model vendors are accelerating penetration into the control layer, threats are closer than expected. Anthropic's Claude Managed Agents already support multi-Agent parallel orchestration; OpenAI further positions the Codex App explicitly as a "command center for agents," supporting multi-Agent parallelism, automation, and long-cycle background running. This means Codex's multi-Agent control capabilities within the software engineering boundary are relatively mature, even surpassing Hermes locally in some parts, no longer a simple "underlying executor."

Hermes temporarily possesses personal control plane advantages across channels, models, and projects; but Codex already possesses strong task ownership and multi-Agent management capabilities within the software engineering boundary, competitive advantages within this boundary may be stronger than Hermes. The core competition question is: Can Hermes, before model vendors, precipitate user project states, approval rules, Skills, Memory, and cross-Agent workflows at its own layer, forming assets users are unwilling to migrate? Or will it ultimately be absorbed as a standard function by model native products?

Internet Big Factory Agent Route Selection

Discussing big factory strategies to cope with the personal Agent wave requires first clarifying their product boundaries: personal resident Agent hosting (such as Tencent QClaw, ByteDance ArkClaw) versus office/enterprise general work Agents (such as WorkBuddy, Trae) have distinctly different positions:

  • Big Factory Claw Route: Lowering thresholds through one-click deployment, preset templates, and local ecosystem access. But the deep chasm lies in untrustworthy platform incentives: No matter how many external models are supported, users naturally believe its ultimate purpose is to divert traffic to its own cloud and model systems.
  • Hermes Runtime Integration: ByteDance ArkClaw and Tencent Cloud have officially accessed Hermes Agent as an optional plugin or exclusive template into their cloud consoles, establishing a clear multi-Runtime strategy: Big factories retain their own cloud hosting, billing, security, and enterprise-level control bases, while viewing Hermes as a pluggable advanced component, achieving complementary symbiosis between open-source ecosystems and commercial cloud platforms.
  • General Office Agent Transformation: Currently, big factories are shifting core resources from Claw to general office Agent platforms with clear demands, easy acceptance, and direct monetization (such as WorkBuddy). Such tasks can be deeply bound to their own ecosystems like WeChat, DingTalk, Feishu and converted into revenue.

Hermes' Implications for Crypto AI

Web3 did not directly make Hermes a smarter Agent, but allowed Nous to possess a capital structure, organizational method, seed users, and value source different from traditional AI startup companies. Hermes proposes at least one more mature Crypto AI path: Let Crypto become the organization and infrastructure, rather than a product interface users must face.

Hermes has already completed the migration from a Crypto AI research brand to a global open-source Agent product, establishing large-scale attributable inference activities and a clear second mindset—but this mindset is currently still concentrated in the OpenRouter ecosystem and global developer circles, not translating into a comprehensive overtaking of OpenClaw in GitHub Stars or overall community scale. It does not have exclusive technology that OpenClaw cannot replicate, but through precisely taking over high-intensity users and establishing "delegatability" and "self-evolution," it completed a challenger product iteration worth studying.

  • Implication One: Crypto can serve as an "Organizational Operating System," rather than product features: The real value of Web3 can be reflected in capital structure, early high-intensity user pools, and value bases, without forcing exposure as wallet or Token interactions. Achieving "Organizational Layer Crypto-native, Product Layer Crypto-invisible" is an effective strategy balancing innovation drive and user experience.
  • Implication Two: Decentralized infrastructure must anchor demand-side entries to form a closed loop: Pure supply-side distributed training networks (such as DisTrO, Psyche) without real user entries and execution data support are difficult to prove commercial value. Hermes Agent is precisely the key verification for Nous crossing from underlying computing infrastructure to the real demand side.
  • Implication Three: Moats can be built on "Delegation Trust" rather than单纯 "Model Capabilities": Differentiation of personal Agents does not necessarily stem from stronger single execution capabilities, but lies in "whether users dare to deliver long-term responsibilities to it." This soft trust asset is a dimension often overlooked but极具 barriers in Crypto AI projects.
  • Implication Four: The relationship with cloud big factories is not a zero-sum game, but ecological complementarity: Big factories accessing Hermes as an optional Runtime proves open-source Runtime and big factory control planes can coexist. For entrepreneurs, "being integrated" is a viable commercialization path, but risks of core value being intercepted by cloud vendor hosting layers must be guarded against.
  • Implication Five: End-game competition will shift from "Single Execution Capability" to "Task Ownership and Trust Accumulation": The most valuable in the future may not be the strongest model at the execution layer, but the "upper-level control system" that can receive final goals, maintain long-term Context, intelligently schedule professional executors, and let users safely deliver responsibilities.

OpenClaw made "individuals owning Agents" a clear product category; Hermes then advanced "long-term delegating Agents" into a more systematic product direction through persistent state, task recovery, evidence acceptance, multi-model supply, and professional Agent delegation. The real test is: When Claude Code and Codex continue to enhance control capabilities within the software engineering boundary, and big factory cloud platforms make multi-Runtime integration smoother, will users still be willing to give final goals and long-term trust to this open Runtime from a Web3 background—and continue to pay for it.