Written by: Merkle3s Capital

The content of this article is jointly presented by Block Analytics Ltd and Merkle 3s Capital. The information contained in this article is for reference only and does not constitute any investment advice or offer invitation. We do not guarantee the accuracy of the content and are not responsible for any consequences arising from it.

Introduction

Code is dead, logic is eternal.

This statement is not false. By 2026, AI can easily translate JavaScript into Python, rewrite Go into Rust, and erase all traces of code plagiarism. However, the homogeneity at the architectural level cannot be eliminated by rewriting. Just as when a book is translated from Chinese to English, its story structure remains unchanged.

In April 2026, a dispute erupted in the AI open-source community regarding "the first rights of self-evolving agents." The opposing parties were the globally leading open-source AI laboratory Nous Research, which raised $70 million and garnered 85,000 stars on GitHub, and the little-known Chinese small team EvoMap.

This dispute constituted the first practical test of the standards for determining intellectual property rights in the AI era:

When code comparison methods fail, what can we rely on to define acts of plagiarism?

Chapter One: Complete Timeline of Events

From February 1 to April 14, in just 74 days, a panoramic record was made of the technological innovation trajectory of a small Chinese team and the development process of a globally leading laboratory.

Comparison of timeline patterns:

From the public release of core concepts (February 1 to 16) to the launch of the Hermes skill ecosystem (March 12), the time span was between 24 and 39 days.

Nous Research claims that its main repository was created as early as July 2025. However, verification results show that the specific module involved in the "self-evolution" accusation was publicly created on March 9, 2026. This time point lags 36 days behind Evolver.

Chapter Two: Definition and Core of Hermes Agent

How does the top AI laboratory with $70 million in funding transform the "unreviewed model" into a "self-evolving agent operating 24/7"?

Background of Nous Research: Who is creating this agent?

With total funding exceeding $70 million (including $50 million in Series A), and 85,000 stars on GitHub, its most famous product is the Hermes series of large language models. This model emphasizes "unrestricted" characteristics, with VentureBeat noting that it outperformed ChatGPT after "removing content limitations."

The current core direction of the team is to promote Hermes from "model" form to "agent" form.

Core positioning analysis: 24/7 autonomous agent

Unlike programming assistant products like Claude Code and Cursor, the positioning of Hermes Agent is closer to the OpenClaw model. It operates as an autonomous agent running in the server background 24/7:

• Deployment form: server resident, no need for IDE binding
• Evolution mechanism: has a learning loop, capabilities continue to iterate with usage frequency
• Memory architecture: supports persistent storage across sessions
• Platform coverage: compatible with 16 mainstream messaging platforms including Telegram, Discord, Slack, WhatsApp, Signal, WeChat, and iMessage
• Cost structure: a monthly $5 VPS can run it

Storage system: three-level memory architecture

Technical base: relies on SQLite FTS5 extension for support. Compared to the OpenClaw's qmd solution, this extension theoretically can achieve more efficient retrieval performance. This architectural design reflects the concept of "digital sovereignty." All memory data and execution logic are retained in the user's local or private environment. This design aligns with GDPR compliance frameworks.

Tool ecosystem: contains 47 built-in tools, follows Anthropic standards

• Function coverage: web search, browser automation, terminal execution, file editing
• Capability extension: image analysis/generation, speech synthesis, code execution sandbox
• Sub-agent scheduling and scheduled task (Cron) mechanism

Hermes Agent adopts the Agent Skills open standard released by Anthropic in December 2025. This standard marks the transition of AI agents from monolithic intelligence to a modular specialized knowledge pattern.

According to industry survey data from the end of 2025, the current vulnerability rate of the skill system is 26.1%. The reason is that systems allow the execution of arbitrary scripts. Hermes prevents the risk of sensitive information leakage through the agent/redact.py module.

Chapter Three: Technical Definition of Evolver

GEP protocol: an "AI Agent Evolution Middleware" program open-sourced 36 days earlier than Hermes.

The main entity of Evolver's development is the EvoMap team (AutoGame Limited). Its current positioning is to serve as "middleware that gives any AI agent the ability to evolve."

Core concept analysis: GEP protocol

GEP (Genome Evolution Protocol): standardized evolution process. Its mechanism is similar to biological gene expression:

• Gene: reusable evolution assets extracted from run logs.
• Capsule: higher-order evolution unit encapsulating multiple related genes.
• Fully audited evolution events: every iteration is traceable

Ten-step evolution closed-loop mechanism

1. Asset file integrity verification 2. Three-level signal extraction (regular matching, keyword recognition, and LLM parsing) 3. Memory map retrieval 4. Dual-track screening of genes and capsules 5. Mutant construction 6. Personality template selection 7. GEP prompt assembly 8. Prompt file writing 9. State solidification and archiving 10. Reflection evaluation execution

Evolver's version iterations are rapid, with 136 versions released to date.

Chapter Four: Focus of Controversy. Accusations of Hermes Plagiarizing Evolver

First observe the architecture, then discern the source code. Is this a coincidence, or is it solid evidence of AI washing code?

4.1 Comparison of Four Core Modules

Module One: Experience-driven reusable asset generation closed loop

Module Two: Three-tier memory architecture system

Module Three: Periodic reflection evaluation mechanism

Module Four: Skill discovery mechanism and on-demand loading scheme

4.2 Analysis of Mapping Relationships of Source Code Modules

EvoMap commentary: In Hermes, there are source code files corresponding to each core module of Evolver. These files are functionally equivalent.

4.3 Terminology System Analysis: Systematic Replacement at the Concept Level

EvoMap performed full-text searches on two code repositories for Hermes.

The search results did not find direct code residue traces.

This situation aligns with typical characteristics of AI cross-language rewriting. In the process of AI architecture rewriting, characteristic strings of the original project are usually not retained. However, architectural homogeneity cannot be eliminated by rewriting operations.

Chapter Five: Nous Research's Response Strategy and Current Public Relations Crisis

"Delete your account." Why does a laboratory valued at $70 million appear so amateurish in its crisis public relations tactics?

5.1 Denial Statement from Teknium

Teknium, co-founder of Nous Research and head of Post-Training, responded on the X platform:

I have never heard of this team before. Today is the first time...

5.2 Official Account "Delete your account" Comment Incident

The official Nous Research account first replied to EvoMap, then deleted that reply and blocked the founder of EvoMap.

This public relations strategy has been widely interpreted in the open-source community as arrogant and lacking technical sincerity. For a laboratory that has raised tens of millions of dollars and is trying to build a global influence, such handling has not only failed to quell the controversy but has further deepened external doubts about its intellectual property transparency.

5.3 Academic Defense Path: Current Status of GEPA Paper

Core defense argument put forward by supporters of Hermes: its inspiration comes from the GEPA paper, not GEP.

GEPA (Genetic-Pareto) is a paper submitted by the Stanford/Berkeley research team to ICLR 2026 titled "GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning." Its core logic is as follows:

• In traditional reinforcement learning paradigms, execution trajectories are compressed into simple numeric scores
• GEPA retains the complete execution path, covering reasoning processes, tool calls, and final outputs
• The humane realization path for LLM error log diagnosis capabilities

However, there is a significant divergence between GEPA and GEP in essential dimensions. GEPA belongs to the category of prompt optimization frameworks, while GEP is a complete agent evolution protocol encompassing a three-tier memory architecture, experience assetization, and periodic reflection mechanisms. The self-evolution system of Hermes is architecturally closer to the GEP paradigm than to the GEPA path.

Chapter Six: AI Codewashing. A New Problem in Contemporary Intellectual Property Landscape

This is the core argument of the entire article, also reflecting the deeper meaning of the title.

6.1 Concept Analysis of AI Codewashing

By 2026, AI technology can easily achieve cross-language "translation" of project code:

• Code migration from JavaScript to Python
• Syntax conversion from Go to Rust
• Complete elimination of traces of plagiarism at the code level
• Yet, the structural homogeneity characteristics still cannot be eradicated.

Traditional code comparison (diff) methods fail in this scenario.

6.2 How to Determine "Structural Plagiarism"?

When code comparison fails, we must resort to architectural analysis:

6.3 Legal and Moral Distinctions

• Legal aspect: "Reference design paradigms" do not constitute infringement; the MIT agreement allows for forking
• Moral dimension: The operation mode of "zero citations + blacklisting" lacks decorum

Ironically, Hermes has cited Anthropic's Agent Skills standard, which in itself reflects respect for academic norms. However, it has not mentioned EvoMap at all.

Chapter Eight: Industry Revelations

Although there are no winners in this controversy, the entire AI agent ecosystem benefits from it.

7.1 Self-Evolving Agents Become Industry Consensus

The correctness of the direction of "Self-Evolving Agents" no longer needs to consider the order of precedence. EvoMap has validated the market demand for this concept with data showing over 1,800 stars earned within ten minutes. Nous Research further confirms that leading laboratories also view this as a certain evolution path.

The future technological evolution pattern of agent products will fully trend towards the capability dimension of "self-evolution."

7.2 Strategic Choice of Open Source Licenses

The Evolver project initially adopted the MIT license. This license has the highest leniency but the weakest protection. After encountering "referencing" (essentially plagiarism), it switched to the GPL-3.0 license. Although the new protocol has stricter constraints, the timing of the adjustment is too late.

Industry revelation: The building of core technological barriers should adopt licensing schemes with stronger constraints (such as AGPL-3.0, BSL, etc.).

7.3 "AI Codewashing" Becomes a New Threat

This presents a new proposition for the open-source community: how to prove "architectural plagiarism after being rewritten by AI"?

Traditional code comparison mechanisms have already failed. It is urgent to establish new assessment standards:

• Analysis of architectural homogeneity
• Traceability review of development timelines
• Verification of citation integrity

7.4 The Eternal Game Between Small Teams and Large Laboratories

EvoMap's innovative efforts over several months are facing the reality that Nous Research, backed by strong resources and industry influence, can replicate them in a short time and capture more market noise.

This script is neither the first of its kind nor will it be the last act. The surpassing of Linux over Minix, the competition between Android and iOS—historical patterns are continuously replayed.

A survival revelation for small teams: The open-source strategy is indeed important, but "open source" is not synonymous with "making free dresses for competitors." Careful selection of protocol terms, meticulous cultivation of community ecology, and carefully constructed brand narratives are all indispensable.

7.5 The Prelude to the "Operating System" War of AI Agents Has Begun

Hermes Agent, Evolver, OpenClaw, Claude Code, and other forces are competing for the same strategic high ground. Their objective is the ecological niche of the "operating system" for AI agents.

The ultimate winner of this contest will hold the core access point of the AI Agent era.

Conclusion

Code is dead, logic is eternal.

In the era of AI cross-language code rewriting, traditional plagiarism determination standards have become ineffective. However, architectural homogeneity remains. Just as translation cannot change the narrative structure of a book, the homogeneity of technical architecture is also difficult to erase.

There are no winners in this dispute:

• EvoMap gained public sympathy and traffic exposure, but the "delete account" response indicates that large laboratories do not see small teams as significant.
• Nous Research gained product visibility and topical heat, but its foundation of community trust was damaged.
• The real beneficiaries are the entire AI Agent ecosystem. This debate made more people aware of the conceptual value of "self-evolving agents."

Regardless of the order of precedence, one thing is certain: the era of AI agent evolution has arrived. At the same time, the new threat of "AI codewashing" has also emerged.

Data sources: X/Twitter (@autogame_17, @Teknium1, @lxfater, @NousResearch), EvoMap technical comparison report, Hermes Agent official documentation, Evolver GitHub repository, GEPA paper (Stanford/Berkeley, ICLR 2026), Anthropic Agent skills standard, EU AI Act compliance documents.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。