Twenty years ago a friend of mine, Andrew, told me about a future where software wouldn’t just be tools we used. It would be agents who worked on our behalf. For most of that time, the idea felt distant. Then this year I built one.

It runs on a small server in my house, talks to me on Telegram, reads my sleep data from Oura, and programs my weight training workouts automatically. And building it completely changed how I think about the future of software.

We are moving from software we open to software we talk to.

Agents are a vision of technology that predates my own experience in tech. So last year when I played around with ChatGPT’s agent functionality, I was impressed with what it could do. When I first heard about OpenClaw in January, I knew this was big. On the advice of another friend Immad, I bought a dedicated box and started running my own home server.

Its been a ride. Learning Linux. Learning how to host my own server. Dealing with the almost daily updates to OpenClaw that sometimes resulted in me deleting data, creating conflicting installations, and starting over again.

Even though I got it set up fairly quickly so I could talk to it through the terminal, access it remotely through the web, and connect it to WhatsApp and Telegram, I spent weeks doing infrastructure work. Security hardening. Setup experiments. Constant tinkering. (Some of that experimentation is exactly what caused the problems when updates arrived.) But eventually I have it now stabilised. And it’s gone from fun to integrating it into my life.

Truth be told, this is the most fun I’ve had with technology in years. It’s hard to explain everything I’m seeing right now, but the short version is this: this is as big as some of the biggest technology shifts before it. I’ll explain one thing I built, which might give you some insight.

But first, some background.

Context

I’ve been weight training since I was 18. But it was only in my mid-30s that I started taking it seriously. A combination of a high quality mobile app, my rugby friends, and some good trainers (like Hans, who’s the best trainer in SF and a new friend JT who’s trained with the world’s best) shifted weight training from something social to pushing towards my physical limits.

Which means I take certain exercises very seriously. For example, the deadlift. A few years ago, I achieve my goal of 500 lbs (230 kg) which Hans put me on. Six months ago  I was doing it for several reps. I’ve probably put on around 20 lbs in the last five years, and it’s mostly muscle. I share this because reaching that level requires smart training. Intelligence as much as discipline. Diet as much as knowing when to stop.

For example, JT taught me to never skip workouts, but reduce their intensity if I’ve had a bad night’s sleep. If my warmups aren’t moving well, I adjust the weight. Because I’m lifting heavy loads, I’m very precise with progression. I log every workout and closely track my stats. But it requires work. I need to interpret my Oura sleep data, track workouts in my training app, and research what programs I should run.

I can already use online LLMs to design programs, send them videos, and get feedback. It’s amazing what has happened in just the last year. But the manual effort is still there. As recently as two years ago, I was thinking about building an app like this to scratch this itch I had.

I didn’t, and thank God. Because my agent just did

Boris

I suggested to the lead author of a training app I use, an open-source weight training program, that he introduce an API. He released it this week. Separately, I integrated my system with the Oura API so I now receive my key sleep and recovery metrics automatically each day. Then I created a strength coach, called Boris, that I’ve instructed to talk to me in broken English via Telegram.

Inside the chat window we discuss my goals and it generates my training program. It can now update the training app almost immediately. If I’m at the gym and I tell it my warmups felt heavy, it updates the program in real time so that when I refresh the app, the new exercise or weight is already there.  It also takes into account my HRV and heart rate from the previous night and adjusts the programming automatically. All I have to do is confirm the weight I lifted.  Sometimes, I sent him videos to assess my technique and that guides what I do next.

This might not sound that impressive, but it is when you consider that just a few years ago this simply wasn’t possible.

Fifteen years ago this would have required building a full software product. Tracking heart rate and HRV, generating personalized programming, and dynamically updating workouts based on real-time feedback and contextual data is a complex system.

That kind of application would easily have cost hundreds of thousands of dollars to build properly. But now I can do it myself. I can connect different systems and sources of knowledge, unify them, and interact with them through a simple chat window on my phone.

The real breakthrough

The breakthrough isn’t the strength coach in my pocket. The breakthrough is that I didn’t build an app.

In the past, software had to know everything in advance. Every feature, every rule, every workflow had to be designed and coded. Now we can connect systems together and let intelligent models reason about the situation. The software doesn’t need to know everything beforehand anymore. It can figure things out.

That changes the economics of building software. Instead of building products, we are starting to build agents. Instead of software that waits for input, we are building software that observes, adapts, and acts.

Personal infrastructure

Something else clicked while building these agents. I didn’t just build tools. I built infrastructure for myself. For most of the history of software, infrastructure was something companies built. Individuals used products that companies created. We saw a move to the cloud and a centralisation of our online lives. But agents change that.

Now individuals can assemble their own systems. Connect APIs. Orchestrate services. Build small pieces of infrastructure that work specifically for them.  Any individual who knows English can build for themselves and that’s a profound shift.

When millions of people start building small pieces of personal infrastructure, the way software evolves will change dramatically.

When software starts acting

Another implication becomes obvious once you start building these systems. For the last thirty years software has mostly been something humans operate. We open apps. We click buttons. We move information between systems.

Agents start to reverse that relationship. Instead of humans coordinating software, software begins coordinating other software. An agent can read data from one system, reason about it, update another system, and then communicate the result back to you. Once that becomes normal, the interface to software stops being the application itself.

The interface becomes a conversation. You stop opening software. You start talking to it.

What happens next

What I built is small. There are a few other agents I have working on other things, and maybe that’s for another time. But the pattern is enormous. Imagine agents that manage your health, finances, travel, work, learning, or investments. Software that understands your context, tracks your goals, and continuously adapts based on new information.

Not dashboards. Not apps. Agents. Software that thinks with you and acts on your behalf.

The web gave us pages. Mobile gave us apps. The cloud gave us APIs. Agents give us software that works for us.

And the technology is improving at an extraordinary pace. Frontier AI models today are already capable of reasoning about complex tasks with the latest knowledge in the world. Open models are catching up quickly and becoming smaller and cheaper. Within a few years, many of these systems will run locally on devices we carry in our pockets.

My head is spinning thinking about how much the world is about to change. Everyone I know who is paying attention to this feels the same way. Things are moving so fast that nobody fully understands where this leads. I’m also not concerned for my buddy Hans and his training business, despite all the doom and gloom about AI replacing humans.

But one thing feels certain. We are entering the age of intelligent personal agents and personal infrastructure. Its going to be a wild ride. And Boris is going to get me to 600lbs on deadlift.