Week 9: Let's build an AI Agent –?Part 1

Hi and welcome back to an exciting week 9 on my journey into a cybernetic life! After raving so much about the potential and capabilities of agentic AI and AI agents, it's time to put it to the test. So, I thought: why not build our own AI agent?

Does this make sense? Probably not—buying is usually a better choice than building—but in this case, it's all about learning. I want to run an experiment: can we create an AI agent that builds its own community on LinkedIn? One that autonomously decides when and what to post and even responds to comments from the community? Let's find out!

But before we start, here are the AI news of the week:

Need-to-know AI News of the week

• xAI, Elon Musk's AI company, has announced Grok-3, a new AI model that has outperformed most benchmarks – even without agentic reasoning. In the Chatbot Arena, where users blind-test AI models, Grok-3 quickly claimed the #1 spot (Leaderboard).
• Meanwhile, OpenAI has rolled out Operator in most regions where ChatGPT is available –?except for the EU, Switzerland, Norway, Liechtenstein, and Iceland ??.
• Not directly AI-related, but still fascinating: Microsoft has unveiled the Majorana 1 quantum chip, the world’s first based on a Topological Core architecture. By leveraging topoconductors, it enables stable, scalable qubits – potentially allowing quantum computers to solve industrial-scale problems in years rather than decades.

What are AI Agents vs. agentic workflows?

In our experiment, we aim to build a true AI agent. To avoid confusion, let's clarify the distinction between agentic workflows and true AI agents, as these terms are often misused.

• Agentic workflows are predefined workflows that prompt LLMs, also multiple times, to process data and branch execution based on mainly predefined steps. While these workflows can automate tasks, they lack true autonomy.
• AI agents do not follow predefined workflows. Instead, they operate based on goals, memory, and tools, deciding for themselves how to achieve their objectives and how many loops are needed. They employ advanced reasoning to devise strategies, execute autonomously, and continuously reflect, adapt, and learn as they progress.

We want to move beyond static workflows and build an AI agent that thinks, adapts, and acts independently.

Let's start with the fundamental idea

Before we start building, it's always a good idea to have a plan. The goal of our experiment is to create a LinkedIn community—entirely managed by an autonomous AI agent. Here's how it could work:

1. Creating a LinkedIn Page: I set up a company page on LinkedIn called "The AI Augmented Human."
2. Registering an LinkedIn App: I register an app that allows the company page to post content on its behalf. This requires LinkedIn’s Community Management API, which comes with an approval process.
3. Building a CustomGPT: I create a CustomGPT, a ChatGPT with additional context. It knows all previous articles and my personal perspective for "The AI Augmented Human." It also understands my typical tone of voice, acting as my digital twin.
4. Fine-tune the CustomGPT on LinkedIn Best Practices It learns how to grow a LinkedIn community, including how often to post, what kind of content to share, and how to generate engagement.
5. Formulate the Goal: The AI's objective is to gain as many followers as possible and maximize engagement—purely through high-quality content.
6. Providing the AI with access to 4 key Tools: The LinkedIn API to fetch current stats and post content. The Google Gemini API to access up-to-date news. The Bing Image Search API to find relevant images for posts. A memory function so the agent can remember information and use it as context in future prompts, allowing it to learn and improve over time.

After completing these steps, I set up a loop that periodically prompts the LLM (for example, every 5 to 30 minutes), allowing it to determine if any actions are necessary to achieve the goal and to execute them accordingly.

Let's build the AI Agent using AI

Since we're on the topic, we also want to use AI to develop our AI agent. As I mentioned last week, I hardly write code myself anymore because it's so convenient. I mostly use the chat in my IDE to tell CoPilot what to do. Even when only a single line needs changing, I often prefer writing a prompt instead of editing it myself – probably out of habit and curiosity. ??

My goal is to develop this AI agent entirely through prompts, with minimal hands-on coding. For this, I use VSCode with GitHub Co-Pilot Chat, which runs GPT-4 in the background and generates solid, high-quality code.

We are ready to build

Next week, we will start developing the AI agent. Until then, I will request access to the Community Management API and set up the GitHub repository. So stay tuned –?exciting things are coming next week!

Already 1'500 subscribers! Help me grow the community

Do you like my newsletter? Then please help me spread the word and grow the community. My goal is to reach 2'000 subscribers before the snow is entirely melted away!

?? Re-post this article

?? Invite your colleagues to subscribe

?? Comment your wishes for future editions!

I wish you a successful agentic week!

?? Christian