Agentic Artificial Intelligence (AI) is rapidly gaining prominence, offering capabilities beyond simple chatbots by enabling agents to perform tasks. However, deploying and managing these agents can be complex and resource-intensive, requiring continuous operation, model subscription costs, and robust security measures. Abacus AI's Supercomputer aims to simplify this process by providing a secure, AI-native managed cloud server. This article details the setup of the Hermes agent on the Abacus AI Supercomputer, enabling 24/7 operation and accessibility from various devices.

Understanding Abacus AI Supercomputer

Abacus AI Supercomputer is presented as an integrated platform for building, deploying, and hosting applications and agents. Users can opt for a built-in agent or deploy custom agents like Open Clo or Hermes. A key feature is the flexibility to utilize any Large Language Model (LLM). This can be achieved by connecting existing model subscriptions or leveraging Route LLM, which intelligently selects the optimal model for a given task from a range of frontier models.

The Supercomputer environment operates on Ubuntu, functioning as a virtual private server (VPS) separate from the user's local machine. This separation enhances security by preventing direct access to personal files and mitigating risks associated with an agent operating autonomously.

Deploying the Hermes Agent

The deployment process for the Hermes agent on Abacus AI Supercomputer is designed for simplicity. After selecting the Hermes agent, a crucial setting is the "always-on" toggle. Enabling this ensures the agent remains active without cold starts or interruptions, facilitating continuous operation. This toggle can be deactivated at any time.

Upon clicking "install," the system automatically configures the necessary components, including a persistent cloud virtual machine with full terminal access, web browsing capabilities, and AI agent functionalities. The setup does not require manual configuration of LLMs, as it can connect to Abacus AI's existing infrastructure for enhanced web search or integrate with user-provided APIs. The entire process, from initiation to a running Hermes agent, is described as taking only a few seconds with a single click.

The Supercomputer platform supports the deployment of multiple agents, such as Hermes and Open Claw, within a single virtual machine, allowing them to operate independently of the user's personal computer. Users also have the option to configure the server to "shut down when inactive" to conserve credits, in contrast to the "always-on" setting.

Interacting with the Hermes Agent

Hermes is characterized as an agent that handles actions and remembers user interactions and requests, improving its performance over time. The agent's dashboard provides configuration options, including chat history, file access, model selection, automation controls, skills, and plugins.

For initial interaction, users can select "Route LLM" as the primary model. The agent can be personalized by providing information about the user, such as name, response preferences, and learning interests. This information is stored using an "Update memory" tool, enabling the agent to recall these details in subsequent interactions. For instance, after updating memory, a new chat session can query "Who am I?" and the agent will retrieve the previously stored information.

Automating Tasks with Skills and Messaging Gateways

The Hermes agent can be configured with automated tasks, such as a daily briefing report. This involves creating a "skill" that researches specific topics, like uses of Agentic AI, and generates a summary with links. This skill can then be integrated into an "automation" (a cron job) and connected to messaging channels for direct delivery.

Hermes supports numerous messaging channels, including Telegram, Discord, Slack, WhatsApp, Signal, and email. For integration with Telegram, users can obtain a Telegram bot token. The setup can be completed manually or via a QR code scan, which streamlines the process by directly linking to Telegram for bot creation. Once connected, the agent can send messages to the user's phone, allowing for interaction and control from mobile devices.

The agent's ability to learn and adapt is demonstrated through feedback mechanisms. If the daily briefing content is not aligned with user preferences, feedback can be provided, and the agent will update the skill accordingly. This iterative refinement process allows the agent to generate more relevant and focused outputs, such as a "consumer-focused" briefing.

The platform also allows for the creation of "Hermes topics" within Telegram, enabling multiple distinct conversations and automations to be managed under a single bot, each with its own set of automations or news feeds.

Model Configuration and Cost Management

Abacus AI Supercomputer allows for granular control over LLM usage. While "Route LLM" is the default for general tasks, users can configure specific LLMs for auxiliary tasks. For example, a particular model can be designated for vision or image analysis to optimize for cost or quality.

The platform also supports a "mixture of agents" configuration, where multiple smaller models collaborate on research, with a more powerful model synthesizing the results. This approach can enhance output quality but may incur higher costs.

The Supercomputer utilizes a credit-based system for billing. The cost is one credit every 5 minutes of operation. For continuous 24/7 operation over a 31-day month, this amounts to approximately 9,000 credits. Abacus AI offers two pricing tiers:

• Basic Tier: $10 per month, providing 20,000 credits. This tier does not include access to Supercomputer.
• Pro Tier: $20 per month, providing 30,000 credits. This tier includes access to Supercomputer.

Running an agent 24/7 on the Pro tier would consume less than one-third of the monthly credits, with the remainder available for LLM usage. This model allows users to leverage powerful frontier models without the need for their own hardware, which can be riskier and more expensive.

Advanced Features and Future Development

The Abacus Supercomputer hosts a dashboard version of Hermes. Advanced features include connecting to local models, restricting agent access within Docker containers, integrating with LLM Wikis, Git, external devices, and Obsidian. The platform also supports external memory providers to enhance the agent's memory capabilities, with further details planned for future discussions.

Users can stop the Supercomputer at any time to halt credit consumption. However, it is noted that if the Supercomputer is not running when an automation is scheduled, the automated task will not be executed. This allows for strategic shutdown periods to manage costs, with the understanding that scheduled automations will be missed during these downtimes.

The article concludes by highlighting the benefits of a 24/7 cloud-based AI agent accessible from various devices, capable of automations and integration with other tools. The author expresses satisfaction with Hermes's ability to improve over time and its seamless integration with the Abacus Supercomputer.