From Shrimp to Horses: The AI Agent Landscape’s Species Migration

1. When "Shrimp" Meets "Hermes"

In early 2026 developers who said they were "raising shrimp" were referring to OpenClaw, an AI Agent nicknamed "little shrimp" for its extensive plugin ecosystem and deep IDE integration. It runs inside VS Code, handling code writing, testing, and PR creation.

On February 25 a project called Hermes Agent appeared on GitHub. Chinese developers likened its name to the luxury brand Hermès, calling it "horse". Within seven weeks the repository earned 84.4 k stars and topped the OpenRouter compute leaderboard, shifting attention from shrimp to horse.

Transitioning from shrimp to horse reflects a fundamental shift in developers' expectations for AI Agents.

2. What Makes Hermes Different?

Hermes Agent originates from the U.S. Nous Research lab, led by Jeffrey Quesnelle, Karan Malhotra, and Teknium, who previously built trust with the Hermes series of fine‑tuned models. Hermes challenges the industry assumption that an Agent should start from a blank slate each time.

Most agents, including OpenClaw, are stateless: conversation memory is cleared after each session, so repeated instructions are forgotten. Hermes asserts that it should remember.

Three‑Layer "Memory Palace"

Hermes implements a four‑layer memory system:

Session memory : context of the current conversation, a basic capability of all agents.

Scenario memory : cross‑session interaction history, remembering tasks performed last week.

Semantic memory : generalized knowledge distilled from multiple interactions, e.g., "the user's project is written in Go and deployed on Tencent Cloud".

Skill memory : the killer feature that extracts successful complex tasks into reusable Skill files stored in ~/.hermes/skills.

After each task Hermes evaluates whether the execution is worth remembering—if a tool is invoked more than five times, if errors are self‑repaired, or if the user corrects the outcome—Hermes writes a Skill. Future similar problems invoke the Skill directly, bypassing re‑reasoning.

In plain terms: OpenClaw is a diligent but forgetful intern; Hermes is an experienced employee that gets smarter with use.

3. WeChat Integration – The External Breakthrough

On April 11 Hermes announced native WeChat support via Tencent’s official iLink Bot API, not a third‑party hack. Users run the command hermes gateway setup, scan a QR code, and turn their WeChat client into the Agent’s command interface.

"In China, if you can master WeChat, you’ll take off."

Chinese developers use Agents for more than coding—troubleshooting, monitoring, data analysis, and daily reports—often from mobile devices. With 24/7 WeChat availability, the Agent evolves from a development tool into a digital employee.

4. Cloud Vendors Compete to "Raise the Horse"

Within three days of the WeChat news, both Tencent Cloud and Alibaba Cloud released official support for Hermes Agent. Tencent Cloud offered a dedicated Lighthouse image for one‑click deployment; Alibaba Cloud followed with a community‑edition solution via Compute Nest.

Two major cloud providers backing the same open‑source project in the same week is unprecedented in the AI Agent space.

The commercial logic is clear:

Hermes runs independently of local devices, making it ideal for long‑running cloud services.

Isolation on cloud servers improves security.

7 × 24 online capability drives continuous compute consumption and revenue.

The self‑evolution feature yields higher user stickiness than one‑off tools.

Hermes also ships the hermes claw migrate command, enabling one‑click migration of OpenClaw configurations, memories, and skills, positioning Hermes as the easiest migration path for product managers.

5. Shrimp vs. Horse – An Architect’s Decision Matrix

Technical leaders face a single question: which platform to back?

Key dimensions compare the two:

Core positioning : OpenClaw – IDE coding assistant; Hermes – full‑scenario autonomous agent.

Memory capability : OpenClaw – session‑level, lost when closed; Hermes – four‑layer persistent memory that improves over time.

Interaction entry : OpenClaw – inside the IDE; Hermes – WeChat, enterprise WeChat, QQ, Feishu, terminal.

Execution mode : OpenClaw – follows the developer online; Hermes – runs independently in the cloud 24/7.

Ecosystem maturity : OpenClaw – rich plugin ecosystem, large community; Hermes – rapid growth, still early.

Applicable scenarios : OpenClaw – individuals or teams focused on coding; Hermes – organizations needing cross‑scenario automation.

If you need a smart coding companion, OpenClaw remains competent. If you aim to make an Agent part of your team’s infrastructure, Hermes may be the better investment.

6. The Direction of the Tide

Hermes’s breakout reflects a deeper industry shift: the consensus is moving from "what an Agent can do" to "how an Agent can improve over time".

Many agents over the past year delivered impressive demos but failed in real deployments because their capabilities were static—limited by the model’s ceiling and prompt engineering.

Hermes answers a critical question with a closed‑loop learning system: an Agent can be a growing system, not just a tool.

When developers start caring about whether an Agent can remember and get stronger, stateless frameworks become as obsolete as pocket calculators in the smartphone era.

Thus the migration from shrimp to horse is not merely a trend swing; it signals a collective upgrade in expectations for AI Agents. The next question is no longer "whether to use an Agent" but "whether your Agent still remembers who you are".