The AI Diary: Build Your Own AI Agent

Feb 18, 2026

Today, I’m going to be more practical. Instead of just talking about AI agents, we’re going to build one. So roll up your sleeves, open your computer, and follow the steps below.

I’ll try to make this manual as simple as possible.

Why?

Reading about AI agents is one thing, but trying them yourself is another. It will give you a real understanding of what they can do and how you can use them.

Terminology

Before we dive deeper, I’ll try to explain a few simple concepts.

Large Language Model (LLM) – Imagine this as a computer that can receive a question from you and, based on that question, builds a sequence of words forming sentences and paragraphs, which it then returns to you as a response. This is basically what happens when you type to ChatGPT, Gemini, Grok, etc.

In other words, we can think of an LLM as a machine that tries to predict the response to give you based on the information you’ve provided before.

An LLM is not conscious; it has no soul. It will provide a response if you ask a question. If you don’t ask anything, it will simply wait.

Tool Calling – LLMs are not just able to generate text; they can also interact with external tools. For example, if you ask an LLM, “What is the weather right now?”, it can perform what’s called tool calling — meaning it uses a tool to check the current weather forecast published somewhere on the Internet.

Another example of tool calling is a tool that allows the LLM to read your emails or send messages on WhatsApp, similar to how humans use tools to interact with the real world.

OpenClaw – Simply put, think of it as a few thousand lines of code (a piece of software) that performs several distinctive functions:

• It has its own memory, stored as simple files containing data.
• It connects to an LLM and sends these memory files to it. As a result, as we mentioned earlier, the LLM generates a response, which OpenClaw documents in its memory files.
• This process runs in an endless loop. OpenClaw constantly calls the LLM, generating and storing new data, allowing it to evolve over time.

For example, if we store a task in these files like: “Check my unread emails, and if I receive an email from person X, notify me on WhatsApp”, OpenClaw will send this task to the LLM. The LLM will use tool calling to check your email inbox, and if there are unread messages from person X, it will send you a message on WhatsApp.

By the way, this endless loop is managed by a feature called the “heartbeat”. OpenClaw’s heartbeat is programmatically configured to activate every 30 minutes to perform tasks, then go to sleep, and wake up again when the next heartbeat occurs.

Server – This is the computer where we install OpenClaw. It is advisable to rent this server rather than use your own computer, since we cannot be 100% sure that your personal data will remain safe if OpenClaw is installed locally.

Technologies We Will Use

LLM

We will use the Claude platform. Steps:

1. Open this link: ​https://platform.claude.com/​
2. Register an account.
3. As of writing this article, go to Settings → Billing to add your card and top up your account with, say, $10–$20.
4. From the left sidebar menu, choose API Keys. Click + Create Key. Store the key somewhere safe on your local machine.

We will use this key later by providing it to OpenClaw, as it is the key OpenClaw uses to access Claude and the LLM they provide.

Server

To host our OpenClaw, we will use the platform exe.xyz for the following reasons:

• For a small monthly fee, you get the ability to build multiple servers, each of which can potentially host its own OpenClaw. At the time of writing, the price is $20 for 25 virtual servers (also known as virtual machines), which comes out to just $0.80 per machine.
• It is super easy to use.

Steps:

1. Open ​exe.dev​ and sign up.
2. Purchase the monthly subscription.
3. Once ready, open this link: ​exe.new/openclaw​.
4. Give your server a name or keep the default name they propose.
5. In the text field containing ANTHROPIC_API_KEY=<fill-this-in>, replace <fill-this-in> with the key you generated earlier.
6. Press Create VM. Your server, with OpenClaw hosted and running, should be ready in a few minutes.

Well, we’re ready. OpenClaw is running with its heartbeat and connected to the LLM. But we still have some work to do — we need to configure a few more things.

The Downside of AI Agents

In my opinion, the biggest downside of AI agents right now is that setting them up isn’t easy for someone without a background in software. The learning curve is steep and it takes time. However, I suspect that gap will diminish soon, and setting up an agent — or even a team of agents — will become so simple that even a child with no prior knowledge could do it.

But before that, stay with me as we dive into a few more technical aspects of setting up OpenClaw.

Set Up OpenClaw

Open exe.xyz in your web browser, and from the main menu, choose VMs. A page with a list of virtual machines will open. Click on the one you just created — this is where OpenClaw lives.

Next, you’ll see a list of buttons, one of which is called Open Shelley Agent. Clicking this will open a new tab where you can interact with an AI agent provided by exe.xyz, which will assist you in setting up your OpenClaw.

We will need to ask the AI agent to configure a few things for us.

First Step

The first time you open the Shelley Agent, ask it the following:

This instruction may look odd, but it is a necessary first step for OpenClaw to run properly.

Model

The API key we generated above connects your OpenClaw to Claude, which, as we mentioned earlier, gives you access to multiple types of LLMs (Large Language Models).

Depending on your goals, you may want to choose different models. Some models are cheaper because they are less powerful compared to the more advanced (and more expensive) models.

For our tests here, I recommend starting with Claude Haiku 4.5. The reason is that Haiku 4.5 does a decent job with the tasks you give it and can also perform tool calling.

You can check the full list of models here: ​platform.claude.com/docs/en/about-claude/pricing​

Now, inside the Shelley Agent chat, ask the agent the following:

Theoretically, the Shelley Agent should configure everything for you. However, since this is not a software program with a predetermined sequence of actions, but rather an interaction with an AI agent, there is a certain level of unpredictability. Therefore, you may encounter situations that are beyond the scope of this article and will need to resolve them on your own.

If you need help, I’m happy to assist. You can always contact me via the contact form on this page: ​www.paveltashev.com/contact/​

Connect to a Communication Channel

To speak with OpenClaw, you need to connect it to an application like WhatsApp, Slack, Telegram, Discord, or another messaging platform. This way, OpenClaw can act like a “person” who can chat and discuss tasks with you.

In my case, I use Telegram because it is very easy to configure. However, feel free to use any other app that is convenient for you. Open the Shelley Agent again and ask it the following:

The Shelley Agent will provide instructions on how to set up the communication channel. Once it is configured, you will be able to open your app and chat with your own OpenClaw.

What’s Next?

It really depends on what you want to use OpenClaw for, but my advice is to experiment. For example, one of the first things I did with my first OpenClaw (her name is Ashley) was to ask her to check social media every six hours for hot topics related to AI, startups, and entrepreneurship, and to prepare a summary highlighting the pain points people discuss. This simple task saved me time spent on researching and digesting information.

Here is a small experiment you can try. Open your messaging app connected to your AI agent and ask it the following:

The reason this type of tool is so powerful is that it remembers what you discussed before, tracks what it has worked on, adapts, and essentially learns from its own experience.

Tools

Keep in mind that if you want to configure your agent to read your emails or access content from social media, you may need to connect it to the corresponding tool to achieve that. This can be a pitfall, as the process can sometimes be quite technical. In such situations, I advise you to open the Shelley Agent and ask for assistance.

For example, I’ve heard that some people connect their OpenClaw to trade crypto. You could experiment by asking Shelley to help you connect your OpenClaw to Binance. However, I strongly advise you to be extremely careful with this, as delegating trading to a bot without extensive trading experience is very risky.

Hallucinations and Lies

Be careful when choosing an LLM model. For example, if you opt for a cheaper model, you may end up with an LLM that gives you false information. I experimented with an LLM called LLaMA 3 70B. When I asked my agent if it could read social media channels, it confidently told me that it could — and even provided a few post titles. Later, I discovered that this was completely false. When I asked the model why it lied, it apologized, saying it simply wasn’t capable enough.

In the end, I ended up paying for an LLM that misled me. This is why it’s better to choose a slightly more expensive and powerful model to ensure that the tasks you assign will actually be completed.

Cost

There is one major downside to all these AI agents: they cost money. Every time you send a question to the LLM (in our case, Claude), you are charged for each word in that question. The longer the text, the higher the cost. Not only that, but the longer your OpenClaw is “alive”, the larger its memory becomes, which can also increase costs.

For example, my first OpenClaw was connected to LLM Opus 4.6, and I ended up spending nearly $40 in just two days. Doing the math, that’s around $600 per month — and I must emphasize, that was for simple chatting, with no specific tasks completed.

So, be careful and experiment cautiously.

OpenClaw Issues

My current observation — and not just mine, but also that of my colleagues — is that OpenClaw has a fundamental architectural issue. This is why I consider it more of an experimental playground, which requires time before it is really ready for building stable production projects.

As I mentioned earlier, OpenClaw has something called a heartbeat. This is a programmatically configured clock that “wakes up” the agent every 30 minutes to perform its tasks.

On top of that, the agent can configure its own “clocks” to run every hour, every two hours, every five minutes, etc. The technical term for these clocks is CRON jobs.

Additionally, the agent may decide to create a daemon process. A daemon is a program that runs continuously as a background process on a Linux system, performing tasks and providing essential services.

Now, every time the heartbeat, these clocks, and daemons run, they read the memory and call the LLM. This can potentially lead to two major issues:

1. High costs, because these processes send a lot of data to the LLM.
2. Task overlap: if any of these tasks (and there can be many) run simultaneously, the same task may be executed twice or multiple times, leading to unpredictable results.

I say this based on personal experience. One of my OpenClaw instances created a mess in the file system as a result of executing multiple tasks at the same time.

Bottom Line

Install, experiment, explore.