Tech Frontiers · TECH

From 'Lobster' to 'Hermes': AI Agents Shift Gears

Approx. 3,500 words · ~10-minute read

Over the past three months, the hottest topic in developer communities has been taming a 'lobster'—OpenClaw, the open-source Agent framework with 350,000 GitHub stars, once hailed as a universal solution. But in the second week of April, the wind shifted on GitHub. A new framework called Hermes Agent burst onto the scene, amassing 57,000 stars in just two months—and posts declaring 'I ditched OpenClaw for Hermes' began appearing in rapid succession. 57,000 starsposts declaring 'I ditched OpenClaw for Hermes' began appearing in rapid succession.

On the 'lobster' side, security vulnerabilities surfaced, its founder joined OpenAI, and Anthropic adjusted pricing for third-party tool access—three heavy blows landed simultaneously. On the 'Hermes' side, the project went from zero to launch in two months—and released five versions in 18 days.

While you’re still figuring out how to raise your lobster, your neighbor is already debating whether to upgrade to Hermes.What’s the difference? It’s not about which one connects to more platforms—but which one gets smarter the more you use it.

01 | What Is It?

Let’s start with architecture. Hermes Agent has three layers—think of it as a company.

The platform adaptation layer is the front desk.It supports 15+ messaging platforms—including Telegram, Discord, Slack, WhatsApp, Signal, Feishu, DingTalk, WeCom, and more. No matter where users enter, the front desk receives them and translates their input into a unified format for internal processing.

The core Agent logic layer is the brain.A class called AIAgent—roughly 10,000 lines of code—handles understanding user input, deciding next steps, invoking tools, and formatting results. This is Hermes’ heaviest layer.

The execution backend layer is the hands and feet.It connects to external tools via MCP (Model Context Protocol—a standardized interface for AI tools, akin to USB-C for AI). File I/O, API calls, browser automation, database access—all plug in via this standard protocol. The hands and feet aren’t built in; they’re attached via standard interfaces.

02 | A Skill Library That Gets Smarter With Use

This is the part of Hermes most worth unpacking.

Most AI tools operate on a 'ask-and-answer' logic: each conversation starts from scratch. Even if the tool helped you deploy a service last time, it will walk through the same steps again the next time—no memory, no reuse.Hermes’ Skill Library aims to break that cycle.

What Is a Skill?A structured Markdown file stored locally under ~/.hermes/skills/. Each skill documents the full operational path for a complex task—steps, known pitfalls, and verification methods. It’s not a code plugin; it’s an experience document.

How Do Skills Emerge?Two paths. First, creation from experience: when a task involves five or more tool calls and succeeds, Hermes autonomously decides whether to distill the process into a skill. For example, deploying a Node.js service on a server—checking environment, installing dependencies, configuring Nginx, setting up SSL, adjusting firewall rules, verifying connectivity—12 steps total. Upon completion, Hermes compresses those steps into a skill and records the pitfalls encountered along the way.Next time deploying on a similar server? Three steps.

Second, refinement through reuse: when encountering a similar scenario, Hermes compares outcomes. If the new approach performs better, it updates the skill. If new pitfalls emerge, it adds them.

Zero resource overhead at startup.At launch, Hermes loads only the name and one-line description of each skill—just over 3,000 tokens for 40+ skills. Full content loads only when a specific skill is needed. This progressive disclosure avoids overwhelming the AI’s 'work desktop' at boot.

Go deeper: Hermes includes GEPA—the Genetic-Pareto Prompt Evolution engine—an experimental genetic algorithm that automatically evolves skills and prompts. Its academic foundation is an ICLR 2026 oral paper, whose core finding is: 'Reflexive LLM reasoning combined with evolutionary algorithms can outperform reinforcement learning.' GEPA remains experimental—but points toward a future where skills learn not just from human usage, but evolve autonomously.

03 | Three-Tier Memory: It Remembers More Than Just Conversations

You may have used ChatGPT’s 'memory' feature—where what it remembers often isn’t what you intended, and editing or deleting entries is opaque and inaccessible. Hermes takes a radically different approach.

MEMORY.md records hard facts about projects: tech stack, server configuration, pitfalls from prior deployments.

USER.md records soft preferences: your preferred name, desired response length, favorite tools.

SessionDB is a cross-conversation full-text search engine. Built on SQLite with FTS5 (full-text search extension), it lets Hermes retrieve historical context via search, summarize using Gemini Flash, and inject that summary into the current conversation.

Fundamental difference from ChatGPT’s memory: Hermes stores memory as plain Markdown files—on your own machine.You can open them in any editor, read them, edit them, delete them—fully transparent, fully controllable.ChatGPT’s memory is a black box hosted on OpenAI servers—you manage it only indirectly, via conversation.

04 | Positioning: Not a Replacement—A Complement

Those familiar with Agent tools might ask: How does Hermes relate to Claude Code and Cursor?

• Hermes: Your AI digital employee—running on your server, on call 24/7, growing smarter with every use

• Claude Code: The best AI coding partner—understanding your entire codebase

• Cursor: An AI-native IDE—pushing AI integration to the limit inside the editor

They occupy distinct lanes. Claude Code and Cursor solve 'how to write code more efficiently.' Hermes solves 'how to let an AI Agent continuously handle diverse operational tasks.' Coding is just one thing Hermes can do—and frankly,not its strongest suit。

Back to the 'lobster' and 'Hermes' metaphor—the fundamental difference isn’t feature count, but strategic direction. OpenClaw pursues thebreadth-first path: 50+ platform integrations, thousands of community plugins, 'connects to everything.' Hermes pursues thedepth-first path: ~15 platform integrations—but far deeper investment in self-evolution and persistent memory.

The costs of breadth are now concentrating. ClawHub—the OpenClaw community plugin marketplace—was found to contain over 824 malicious skills; security researchers scanned over 130,000 unauthenticated exposed instances; its founder joined OpenAI in February; and in April, Anthropic changed pricing—Claude subscriptions no longer cover third-party Agent calls.

The wider the surface area you connect to, the wider the attack surface becomes.Hermes chooses to deepen understanding within a limited connection surface—making it especially attractive amid rising security debt.

05 | Not a Silver Bullet

Stating limitations clearly matters more than hyping strengths.

73% of tokens are fixed overhead.46% consumed by tool definitions, 27% by system prompts. Only 27% remain for actual dialogue. Nearly three-quarters of the model’s 'work desktop' is occupied by static metadata.

Default security policy is ALLOW-ALL.Out-of-the-box, the Agent is permitted to execute any action without confirmation. Security audits identified 4 Critical + 9 High severity issues. Developers with strong technical fluency can tighten permissions easily; average users must actively manage them.

Coding is not its strength.Serious coding tasks require Claude Code or Cursor. Hermes can write scripts, modify configs, and perform light development—but that’s not its home turf.

Self-assessment leans optimistic.Post-task self-evaluation is almost uniformly positive. If skills are auto-generated from such assessments, low-quality experience may accumulate.

Self-learning is disabled by default.The self-evolution capability described earlier is not enabled out of the box. Manual configuration is required—and official documentation doesn’t clarify how.

One more thing: Reddit users noted that some highly positive review posts originated from newly created accounts with templated content,suggesting possible astroturfing. That doesn’t prove the product is flawed—but it’s reasonable to remain skeptical of overly uniform praise.

06 | A System That Grows With You

Nous Research—the company behind Hermes—is also noteworthy. Originating in 2023 from a Discord community, it began with open-source LLM fine-tuning. It has since raised a seed round of USD 5.2 millionand a Series A of USD 50 million, reaching a $1 billion valuation. Hermes Agent shipped five versions in 18 days; v0.4 included 300 PRs merged in five days—development pace is extremely aggressive.

Hermes has one easily overlooked feature:MCP Server mode(introduced in v0.7.0). Hermes doesn’t just act as a client connecting to external tools—it can expose itself as an MCP server. Using the open ACP (Agent Client Protocol), a Hermes instance can be directly invoked by mainstream editors including Zed, JetBrains, Neovim, Emacs, and Obsidian. It shifts from 'an Agent' to 'an Agent backend.'

But all these features, capabilities, and architectural choices ultimately circle back to one question:What is the real value of an Agent?

For the past year, AI tool competition has revolved around 'connecting more platforms,' 'supporting more models,' and 'integrating more tools.' Where does that quantity game end? For an independent builder, daily usage likely touches only three to five of those 50 platforms.

What you may need isn’t a universal tool—buta system that increasingly understands you。

Next milestone to watch: GEPA’s evolution engine moving from experimental to production-ready. If skills can not only learn from human usage—but also self-evolve and prune low-quality experience—then Agents stop being mere tools and become true digital colleagues.

Robin Compound Notes · Tech Frontiers TECH