Linux AI Agent: Automating SAP HANA Installation

[L_Skorwider]

The New SAP AI Agent

I created an SAP AI agent. Again. This time, I challenged myself with an operating system-level task and specifically came up with the idea of installing the SAP HANA database using an AI Agent. 

If you're already familiar with installing SAP HANA, you know it involves a lot of hands-on work in the terminal. Sure, you could use a graphical interface or automation tools, but for now, think of this as a proof of concept to show what this agent can do. We're aiming for a versatile tool here, not something that's only good with SAP HANA installation, but something that can handle a wide range of applications and tasks right from the Linux terminal.

The Linux Agent

For various tasks, we already know that being able to work in a terminal like a human is more effective than just executing commands, even though the latter might be easier to implement. Don't get me wrong - I've always said that while using human-like interfaces might not be the most efficient way for artificial intelligence to function, it's still essential because not everything can be done differently.

When working in a Linux terminal, the main tools you rely on are your eyes to see what's happening and your fingers to quickly press the keys. And actually, agents can be equipped with the same tools and expected to work in a manner similar to humans. I've equipped the agent with these three tools. Why three not two? To keep things simple:

• We have a tool that lets you input text. It's pretty straightforward and obvious, so there's not much to delve into here.
• The other tool may be slightly less obvious and is used for handling special keys and their combinations. For example, Backspace, Delete, Print Screen, Escape, Ctrl-C, Ctrl-V, and others. It is clear that this is not plain text, and for instance, the ESC key is invaluable when using VI. In fact, not only there.
• And finally, we have a tool to preview the current contents of the terminal window. Something obvious to use during long-running operations when it is not really expected for the agent to press any keys.

OpenAI Agents SDK

In contrast to my SAP GUI AI Agent, where LangChain was used, this time I opted for the OpenAI Agents SDK framework. Why this one and not LangChain? The answer is simple. In such a dynamically changing industry, you have to try different solutions. And since the OpenAI framework doesn't tie me to models from just this company, but is universal, I decided to try it out. I conducted my initial tests specifically on the DeepSeek V3 model. Of course, when choosing other models, we lose some functionality, but the most important part is available.

Model Context Protocol

Lately, there's been quite a buzz about MCP, and for good reason - it's truly revolutionary. The power of agents comes from their tools, but what makes MCP stand out is its standardization and reusability. By designing an MCP server to work with a Linux terminal, I know I'll be able to use it later with another agent or even a regular chat client. Even the OpenAI desktop application will soon support MCP servers.

I'm confident that MCP will have a world-changing impact. We can use them for a range of tasks without always depending on external agents, just by adding tools to our favorite client. Keep in mind, MCP isn't just about tools; it also includes prompts and other resources. It's clear that someone designed it with the future in mind, potentially to replace whole external agents. I get the feeling that developing well-designed MCP servers will become more critical than creating agents. That's why it's essential to get to know this technology now, and that's why I opted for MCP.

The LLM

Although the OpenAI Agents SDK allows me to use virtually any model, I decided to test using GPT 4.1. This model was unveiled to the world just a few days ago. It is, in a sense, the successor to GPT-4o, but it is better at understanding commands and using tools, which is critical for agent applications.

Since my agent is an experimental solution, the choice of a new model was quite obvious. However, we should be aware that it is not a reasoning model and does not belong to the top tier of the best LLMs. It is, however, economical and very fast, which is evident in the agent's operation.

Hardware Environment

If you're familiar with my approach to miniPCs, it won't be a shocker that I picked this platform for testing. Sure, the cloud is super convenient and definitely the future, but running a system at home still has its perks. I used the Firebat AK2 with 16GB of RAM and installed Red Hat Enterprise Linux version 8.8 on it. It's free for developers and supported for SAP HANA.

Reflections on Safety

Safety in artificial intelligence is a crucial aspect that's often overlooked, but it's definitely something we should focus on. See how we are increasingly easily relinquishing control over our computers connected to the network to artificial intelligence. This approach carries some risk because we still don't completely understand how large language models operate. If you're curious about this topic, I strongly recommend checking out my earlier post: Beyond the Black Box: The Illusion of Control.

Here's the ironic part: having so many MCP servers available and being so easy to use can actually create problems. This happens when artificial intelligence is set up by people who don't fully understand the risks involved.

Summary

The video I recorded as part of the Agent tests is short, but be aware that these are just initial attempts. I'm surprised at how well it went the first time. While I haven't tested DeepSeek on an SAP HANA installation, only on simpler tasks, I see that GPT 4.1 handles agent tasks much better.

We're just starting out on this journey, but it's already clear how much potential there is. I see my agents as a proof of concept that shows how current models can handle a wide range of tasks they weren't specifically trained for. This flexibility is exactly what sets the agent-based approach apart from traditional automation. Sure, automating SAP HANA installations is straightforward and efficient. However, in this case, the agent executes a task efficiently even without prior training. This suggests it could tackle many other tasks just as well.

Where is this going to take us? We'll find out. But one thing's for sure - we're living in interesting times.

I invite you to join the discussion and feel free to ask any questions. If there's enough interest, I'd be more than happy to create a follow-up video.