I am Switching to Hermes Agent

Introduction to Hermes AI Agent

Overview of Hermes Agent

• A new generation of AI agents is emerging, with most forgetting what they learn; however, the Hermes agent retains knowledge and improves over time.

• The current AI agent landscape resembles the web wars of 1995, lacking consensus on essential features and functionalities.

Unique Learning Architecture

• Unlike typical stateless frameworks, Hermes assumes that an AI agent learns from interactions, enhancing its value over time.

• Nick shares his experience in implementing AI for companies and emphasizes understanding architectural choices that differentiate stagnating tools from improving ones.

What is the Hermes Framework?

Key Features of Hermes Agent

• Hermes is an open-source self-improving AI agent framework developed by News Research (version 0.6.0), production-ready with strong community support (19,000 GitHub stars).

• It supports over 40 tools including web search, terminal access, file system interaction, browser automation, code execution, image generation, text-to-speech capabilities, and vision processing.

Communication Capabilities

• The framework includes a messaging gateway allowing communication through various platforms like Telegram, Discord, Slack, WhatsApp, Signal, and email.

Learning System in Action

Built-in Learning Mechanism

• Hermes not only runs tests but also learns from them; it’s primarily built on Python (92.5% of the codebase), facilitating integration with existing LLM APIs and training frameworks.

Practical Application Example

• In practice, Hermes can store all conversations and tasks in a searchable format to recall past interactions effectively.

Skill Creation and Improvement

Skill Development Process

• When tasked with analyzing monthly revenue data from multiple sources (e.g., Stripe and HubSpot), Hermes creates reusable skills based on initial complex interactions.

Continuous Refinement

• As users interact more with Hermes over time—like requesting weekly summaries—it refines its processes to improve efficiency for future requests.

Personalization Over Time

Understanding User Preferences

• The agent tracks user-specific patterns such as preferences or working styles to provide tailored suggestions rather than generic responses.

Compounding Improvements

• This personalized learning approach allows Hermes to become increasingly adept at meeting user needs compared to other agents that reset after each session.

Deployment Considerations

Importance of Deployment Environment

• Where you run an agent matters significantly; many frameworks are limited to local machines which cease functioning when turned off.

Understanding Hermes: Infrastructure and Functionality

Deployment Options for Hermes

• Hermes can be installed in various environments including local terminals, Docker containers, SSH on VPS, Singularity for research clusters, and cloud platforms like Modal.

• Modal allows users to pay only for compute seconds when the agent is idle, making it a cost-effective solution at approximately $5 to $10 per month compared to traditional VPS options.

Project-Specific Infrastructure Needs

• Different projects require tailored infrastructure; personal assistants may run locally while production systems might need SSH or Docker.

• Startups may benefit from serverless solutions that do not lock them into specific infrastructures, allowing flexibility in backend choices.

How Hermes Operates Mechanically

• The operation begins with a user request that requires complex tasks rather than simple queries. For example, pulling revenue data from Stripe and analyzing it against HubSpot data.

• Hermes utilizes over 40 tools to execute tasks involving multiple APIs and programming languages (e.g., Python), which may take several minutes during initial runs.

Learning and Skill Development in Hermes

• After completing a task, Hermes records every action taken—this record is termed a "trajectory," which contrasts with other frameworks that discard this information post-task.

• The recorded trajectory is analyzed to determine if it can be packaged as a reusable skill. If so, it creates a function that can be stored locally for future use.

Continuous Improvement of Skills

• When similar requests are made in the future (e.g., generating reports), Hermes leverages previously created skills to enhance efficiency and accuracy.

• Skills evolve over time based on usage patterns; they refine themselves according to changes such as new data sources or formatting requirements.

Limitations of Using Hermes

Windows Compatibility Issues

• Currently, Windows users face significant limitations as full functionality requires WSL2; experimental support is being developed but remains incomplete.

Learning Loops and Framework Comparisons in AI

Understanding Learning Loops

• The learning loop is theoretically appealing but practically limited; it excels at developing skills for complex tasks while struggling with simpler ones.

• Users may find that the generated skills can fail, leading to wasted time rather than efficiency gains, highlighting the need for realistic expectations.

• Security concerns arise when running agents locally, as they can execute terminal commands on the user's machine without adequate safeguards.

Challenges with Local Agents

• Agents can become unresponsive or stuck in loops, despite recent updates aimed at improving this issue; users should be prepared for these recurring challenges.

• Speed is a concern for multi-task workflows; Hermes prioritizes learning over execution speed compared to other frameworks like Crew AI and Langraph.

Use Cases and Limitations of Hermes

• Hermes is ideal for users focused on building intelligent agents that improve over time but may not suit those needing rapid execution or strict security environments.

• For Mac or Linux users comfortable with Docker, Hermes offers a viable solution; however, Windows users might face limitations.

Comparing Hermes and OpenClaw

• OpenClaw represents a different architectural approach compared to Hermes, focusing on messaging across multiple platforms rather than autonomous learning.

• Developed by Peter Steinberger as a side project in 2025, OpenClaw gained immense popularity quickly, surpassing React's GitHub star record within months.

Features of OpenClaw

• OpenClaw serves as a control plane routing AI interactions across various platforms (e.g., WhatsApp, Telegram), boasting a large community ecosystem with numerous skills available.

• Unlike Hermes, OpenClaw does not learn autonomously; its skills require manual installation and configuration without memory retention capabilities.

Architectural Differences Between Frameworks

• While OpenClaw has extensive community support and functionality across business channels from day one, it lacks built-in learning mechanisms found in Hermes.

• Users must manage multiple plugins for memory systems in OpenClaw versus the integrated learning loop present in Hermes' architecture.

AI Model Infrastructure and Features Comparison

Overview of AI Models: Hermes vs. OpenClaw

• The discussion begins with the capabilities of AI models that self-improve during use, highlighting the RL training pipeline as a crucial infrastructure component lacking in OpenClaw.

• Hermes supports 12 platforms including Telegram, Discord, and WhatsApp, narrowing the gap to OpenClaw's 24 platforms; this cross-platform continuity is emphasized by multiple reviewers.

• Hermes offers six different backends (local, Docker, SSH, Singularity, Daytona, modal serverless), allowing for flexible deployment options compared to OpenClaw's reliance on Node.js or Docker.

• A unique feature of Hermes is its built-in migration tool for OpenClaw users to easily transfer their configurations and data with a single command (Hermesclaw migrate).

• Users are encouraged to run both systems concurrently if they find value in both; Hermes is positioned as a personal agent while OpenClaw serves operational needs.

Choosing Between Hermes and OpenClaw

• For businesses needing an agent across multiple channels with extensive community support and tutorials available at any time, OpenClaw is recommended.

• Conversely, if users seek an agent that learns over time and can export training data for custom models while being comfortable with server management, Hermes is suggested as the better option.

Setting Up Hermes

• The setup process for Hermes is described as straightforward; it involves running a simple command from GitHub to install the necessary components.

• The installer automatically checks system requirements and installs dependencies like Python and NodeJS without user intervention.

• Users only need Git installed prior to setup; once initiated, a wizard guides them through configuration steps seamlessly.

Configuration Steps

• During setup, users can input their existing credentials for services like ChatGPT or Anthropic without restrictions unlike in OpenClaw.

• The TTS (Text-to-Speech) integration allows interaction using services like 11 Labs; users can choose their preferred backend such as Docker or local installation during configuration.

Finalizing Setup

• Users are guided through enabling various features such as pseudo support and context compression based on recommendations provided during setup.

• Customization options are available but defaults can be used initially; specific integrations like WhatsApp web hooks require creating new bots within respective platforms.

Setting Up Hermes Agent

Initial Configuration

• The speaker discusses various skills available for the Hermes agent, including web searching, scraping, and browser automation. They choose to leave most options blank initially but will enable some features later.

• The setup process involves skipping certain configurations like smart home settings and image generation for now.

Environment Setup

• To start using Hermes, the speaker types "Hermes," which initializes the environment and provides access to available tools.

• Users can customize their terminal interface by changing colors for better visibility of commands.

Command Interface

• The speaker mentions a variety of commands available within the Hermes agent, including options for different API keys and integrations with services like OpenAI's models.

• There are numerous connections available such as Obsidian, Claude code, Codeex, and others that enhance functionality.

Mobile Access

• To run Hermes on a mobile device, users must start the gateway by typing "Hermes Gateway." This allows Telegram integration for remote access.

• After setting up the gateway, users can interact with their bot via Telegram on their phones.

Deploying Hermes on VPS

VPS Setup Overview

• The speaker explains how to set up Hermes on a Virtual Private Server (VPS), emphasizing security and ease of use without needing to keep a personal computer running.

Complexity in Deployment

• Acknowledges that setting up on a VPS is more complex than local installation but offers assistance if needed.

The Future of AI Agents

Learning vs. Breadth in AI Development

• The discussion shifts towards the evolution of AI agents like Hermes that focus on learning over breadth. It highlights challenges in architecting adaptable agents.

Community Engagement

• Encourages joining an AI community focused on practical implementation challenges rather than theoretical discussions. Over 18,000 members share insights about building AI into businesses.

Conclusion

• The video wraps up with an invitation to subscribe for more content related to AI infrastructure and agent frameworks.