AI agents, compounding and COVID vibes

Why Matt Shumer's not wrong, and why you are about to feel compounding happen in real time

Written by Andre Molnar 2026-02-19

My friend Patrick recently wrote:

People are starting to hear that something is happening out there again, and it's anyone's guess how it's going to change everyone's lives.

Then Matt Shumer wrote his piece, where he does an excellent job articulating what is happening.

I'd like to focus on the why.

Compounding

We humans have a notoriously bad intuition for exponential growth. We're really bad at gauging just how quickly dramatic compounding effects will occur. That's why we can get caught off guard by something like a pandemic even when we see it coming.

All you need to know, though, is that compounding is multiplicative, not additive, and the biggest growth happens toward the end of the curve.

Next month will have more than two units of change from last month. Tomorrow will have more than two units of change compared to yesterday. This exponential curve started sloping up 30 years ago with Big Data, through Machine and Deep Learning, toward modern AI today.

Remember, the biggest change happens toward the end of the curve.

Cambrian explosion

The Cambrian period was a time of rapid diversification of life on Earth. Millions of species emerged in just a few million years.

For comparison, the first multicellular organisms took more than 1 billion years to appear after the first single-celled organisms. It took hundreds of millions of years before adaptations in multicellular organisms led to their first organs. Then tens of millions of years passed before the first animals appeared in the Ediacaran period.

What happened during the Cambrian explosion? It's not what happened then, but what led to it. Hundreds of millions of years of compounding advantageous adaptations that coalesced into the right basic combination of organs, specialized cells, and a nervous system that could adapt to a wider range of environments.

The right basic combination of general-purpose organic stuff that could adapt in novel ways to different environments.

Evolution of agentic systems

What we are seeing today is just the right basic combination of general-purpose computing stuff that's going to diversify rapidly and adapt to every niche.

The shape of this basic stuff is a simple program loop that can interpret both human and computer inputs with the help of a large language model. The program has instructions to remember what it's doing, instructions to find specialized instructions, instructions to execute new instructions in response to its own outputs, and instructions on when to stop.

The instructions in these programs - prompt templates, context managers, memory stores - are the proto-organs of our modern agentic systems that have been evolving in pockets inside research labs, companies, and niche communities since the general availability of the GPT-3 APIs in late 2021. Thousands of software engineering approaches competing for role-specific fitness across the different functions of this program loop. Millions of configurations in search of just-right tool boundaries.

The proto-organisms that have evolved include Claude Code, GitHub Copilot CLI, Opencode, and Openclaw. They all have the same standard shape thanks to large language models with tool-calling capability, open standards such as MCP and Skills, and the ability to execute simple programs on their host computers. Their internals may differ slightly, but they're all evolving from the same basic shape and the same internal boundaries. They all started as software engineering tools, but they're doing much more because this basic shape applies to many domains beyond software engineering.

Proliferation

Not long ago, getting machine learning (ML) systems to do useful work required programming skill coupled with deep mathematical comfort or a team to support bespoke ML toolchains. A couple of months ago, to extend Claude Code to do something useful beside code, required understanding of software engineering. Today you can set a swarm of agents on non programming workflows with nothing more than computer literacy, a security mindset, and an agent runner. ^†

That means what was once a toy accessible only to a few for specific tasks will now be in the hands of many to accomplish anything. The cumulative creativity of humans is about to be unleashed, for better or worse.

Matt Shumer's not wrong

Change is happening. It's happening fast. It is going to get faster.

Everyone needs to be literate about these tools. Everyone needs to start harnessing these tools' power.

What you can do

The barrier to entry is dropping. It's never been easier to automate^🐉 the tedium out of any task or workflow. It's never been easier to learn from other humans and machines. It's never been easier to interact with the technology to accomplish tasks. The technology has never been smarter or more adaptable to niche use cases than it is today. And it's only getting better.

As a result, more people will use these capabilities, and they will make more things. There will be more of everything: more art, more beauty, and more stupidity.

Don't worry.

Remember that you have the ability to direct these tools toward the first two, and use them to combat the last.

^† Basic computer literacy: Understanding of the operating system, user roles, permissions, secrets, and security. ^↩

^🐉 Warning: Never use a skill or share credentials with software you don't trust. There is a good faith effort by the folks behind clawhub to improve security, but there are a lot of bad faith actors out there. ^↩