Author: Merkle3s Capital; Source: X, @Merkle3sCapital

Introduction

Code is dead, logic is eternal.

This is not alarmist. In 2026, AI can easily translate JavaScript into Python, rewrite Go into Rust, and eliminate all traces of plagiarism at the code level. But architectural isomorphism cannot be eliminated by rewriting—just like if you translate a book from Chinese to English, the story structure is still that story structure.

In April 2026, a war broke out in the AI open source community about who invented the self-evolving Agent first. On one side is Nous Research, the world's top open source AI Lab with $70M in financing and 85,000 GitHub Stars; on the other side is a small Chinese team called EvoMap.

This is the first practical test of intellectual property judgment standards in the AI era:

When code comparison fails, what else can we rely on to determine plagiarism?

Chapter 1: Full timeline of events

From February 1st to April 14th, 74 days, a small Chinese team’s technological innovation and the entire process of the world’s top Lab development.

Timeline comparison:

There is a gap of 24-39 days between the core concept disclosure (February 1st-16th) and the Hermes skill ecosystem release (March 12th)

Nous Research claims that its main repository was created in July 2025. But verification showed that the alleged "self-evolution" specific module was publicly created on March 9, 2026 - 36 days later than Evolver.

Chapter 2: What is Hermes Agent

How a top AI Lab with $70M in financing turned "uncensored model" into "7/24 self-evolving agent".

Nous Research: Who is building this Agent?

Financing exceeds $70M (including $50M Series A), GitHub 85,000 Stars. Their most famous product is the Hermes series LLM - a large language model known for its "no censorship". VentureBeat evaluated its "outperform ChatGPT without content restrictions".

What they are doing now is:Changing Hermes from "Model" to "Agent".

Core Positioning: All-Weather Autonomous Agent

Different from "programming assistants" such as Claude Code and Cursor, Hermes Agent is positioned as an autonomous Agent similar to OpenClaw's 24/7 server background running:

• Run mode: Server resident, not IDE bound

• Evolutionary ability: There is a learning cycle, the more you use it, the stronger it becomes

• Memory: Cross-session persistent memory

• Platform: Telegram/Discord/Slack/WhatsApp/Signal/WeChat/iMessage and other 16 messaging platforms

• Cost: $5 VPS ready to run

Three-tier memory system

The underlying support is the SQLite FTS5 extension, which can theoretically achieve faster memory search than OpenClaw's qmd. This design embodies the philosophy of "digital sovereignty" - all memory and execution logic remain local or private to the user, in line with the GDPR compliance narrative.

47 built-in tools and Anthropic standards

• Web search, browser automation, terminal execution, file editing

• Image analysis/generation, speech synthesis, code execution sandbox

• Sub-agent scheduling, scheduled tasks (Cron)

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 modular expertise.

According to an industry survey at the end of 2025, the skill system vulnerability rate is as high as 26.1% - because it allows the execution of arbitrary scripts. Hermes uses the agent/redact.py module to prevent sensitive information leakage.

Chapter 3: What is Evolver

GEP protocol - an "AI Agent evolution middleware" open source 36 days earlier than Hermes.

Evolver was developed by the EvoMap team (AutoGame Limited) and is positioned as "a middleware that equips any AI Agent with evolutionary capabilities."

Core concept: GEP protocol

GEP (Genome Evolution Protocol) - standardized evolutionary process, similar to biological gene expression:

• Gene: reusable evolutionary assets, extracted from run logs

• Capsule: a more advanced evolutionary unit that encapsulates multiple related Genes

• Evolution Event: Every evolution has an audit trail

10-step evolutionary cycle

1. Make sure the asset file exists

2. Three-layer signal extraction (regular/keyword/LLM)

3. Memory map consultation

4. Gene + capsule selection

5. Construct mutations

6. Choose personality

7. Assemble GEP prompt words

8. Write prompt file

9. Write solidification status

10. Reflex Assessment

Evolver has released 136 versions so far, and iterations are extremely fast.

Chapter 4: Focus of Controversy - Accusation that Hermes plagiarized Evolver

Look at the structure first, then look at the distance. Is it a coincidence or AI code laundering?

4.1 Align the four core modules one by one

Module 1: Closed loop of generating reusable assets from experience

Module 2: Three-tier memory system

Module 3: Periodic Reflection Mechanism

Module 4: Skill discovery and on-demand loading

4.2 One-to-one correspondence between source code modules

EvoMap Comments:Each Evolver core module has a functionally equivalent corresponding file in Hermes.

4.3 Terminology analysis: systematic concept replacement

EvoMap performed a full-text search on two Hermes warehouses:

No direct code residue found.

This is consistent with the characteristics of AI cross-language rewriting - when AI rewrites the architecture, it will not retain the characteristic strings of the original project, but the isomorphism at the architectural level cannot be eliminated by rewriting.

Chapter 5: Nous Research’s response and public relations crisis

"Delete your account"——Why is a $70M Lab’s crisis public relations so amateurish?

5.1 Teknium’s Denial

Teknium (co-founder of Nous Research, Head of Post-Training) responded on X:

I have never heard of them. Today I heard about it for the first time...

5.2 "Delete your account" of the official account

The official account of Nous Research replied to EvoMap, thendeleted the reply and blocked the founder of EvoMap.

This PR strategy is widely interpreted in the open source community as "arrogant" and lacking technical sincerity. For a lab that has raised tens of millions of dollars and is trying to build a global presence, this approach has not only failed to quell the controversy, but has only heightened questions about the transparency of its intellectual property.

5.3 Academic Defense: GEPA Paper

The core defense offered by Hermes supporters:Inspired by the GEPA paper, not the GEP.

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

• Traditional reinforcement learning compresses execution trajectories into simple numerical scores

• GEPA preserves the complete execution path (inference, tool calls, output)

• Let LLM read error logs and diagnose problems like a human

But the dimensions of GEPA and GEP are completely different: GEPA is a prompt optimization framework, and GEP is a complete Agent evolution protocol (three-layer memory, experience capitalization, periodic reflection). Hermes' self-evolving system is closer to GEP than GEPA in architecture.

Chapter 6: AI Codewashing—Intellectual Property Problems in the New Era

This is the core argument of this article and the real meaning of the title.

6.1 What is AI Codewashing?

In 2026, AI can easily "translate" a project's code into another language:

• JavaScript → Python

• Go → Rust

• Eliminate all traces of plagiarism at the code level

• But the isomorphism at the architectural level cannot be eliminated

Traditional code comparison (diff) fails in this case.

6.2 How to determine "structural plagiarism"?

When code comparison fails, we can only rely on architectural analysis:

6.3 Law vs Morality

• Legal: "Reference Design Paradigm" does not constitute infringement, and the MIT license allows fork

• Morally: "Zero citation + blocking" is very unseemly

The ultimate irony: Hermes cites Anthropic's Agent Skills standards to illustrate their respect for academic citations. But not a word about EvoMap.

Chapter 7: Industry Enlightenment

There are no winners in this controversy, but the entire AI Agent ecosystem benefits from it.

7.1 Self-evolving Agent becomes consensus

No matter who comes first, the direction of "self-evolving Agent" has been verified. EvoMap proved that there is market demand for the concept (1800+ Stars in 10 minutes), and Nous Research proved that big Labs also think this is the right direction.

In the future, all Agent products will move in the direction of "self-evolution".

7.2 Strategic selection of open source protocols

Evolver initially chose the MIT license—the loosest, but also the least protective. After being "learned", it was changed to GPL-3.0 - more strict, but it was too late.

Lessons: Core technical barriers should use stricter protocols (AGPL-3.0, BSL, etc.).

7.3 "AI Codewashing" is a new threat

This raises a new topic for the open source community: How to prove "AI rewritten architecture plagiarism"?

Traditional code comparison fails and new evaluation standards need to be established:

• Architectural isomorphism analysis

• Development timeline review

• Referential integrity check

7.4 The eternal dilemma of small team vs. big lab

EvoMap took months of hard work to innovate, but Nous Research, with greater resources and influence, can "replicate" and gain greater visibility in a short period of time.

This is not the first case, nor will it be the last. Linux is to Minix, Android is to iOS, history keeps repeating itself.

Advice for small teams: Open source is important, but "open source" does not mean "giving warmth to competitors for free." Protocol selection, community building, and brand narrative are all indispensable.

7.5 "Operating system battle" has just begun

Hermes Agent, Evolver, OpenClaw, Claude Code - they are all vying for the same position:the "operating system" for AI Agents.

Whoever wins will control the entrance to the AI Agent era.

Conclusion

Code is dead, logic is eternal.

In an era when AI can rewrite code across languages, traditional plagiarism determination standards are no longer effective. But architectural isomorphism cannot be eliminated—just as you cannot change the story structure of a book through translation.

There are no winners in this dispute:

• EvoMap has won sympathy and traffic, but the response of "Delete your account" shows that Big Lab does not take small teams seriously

• Nous Research wins product and buzz, but community trust suffers

• What really benefits is the entire AI Agent ecosystem - because this debate has made more people aware of the concept of "self-evolving Agent"

No matter who comes first and who comes last, one thing is certain: The evolutionary era of AI Agent is here. And the new threat of "AI Codewashing" has also arrived.