Intelligence in the Age of AI Agents

The year was 2011. We called AI “machine learning” back then, and I was working to grow my hedge fund I’d recently launched with 3 other partners. We used python to scrape financial statements + pull them from EDGAR and then we performed sentiment analysis alongside more traditional financial analysis. Fund failed, but I fell in love with the leverage tech created and the new capabilities it made possible.

So I pivoted to tech investing.

I saw a lot of deals (2,000+ pitch decks, 800+ pitches) and invested in a dozen companies. Went on diligence visits to see close to 50 companies. I remember one of my first office visits. This company was run by two scary smart programmers out of a space in Brooklyn that smelled permanently of pizza wafting from the slice shop below. Their great innovation was a novel chess engine they were applying to business problems in unique ways. It was brilliant, of course. It could thrash a grandmaster without breaking a sweat, calculating 3 billion moves a second while its server fans hummed a tune of quiet superiority. They were using this mind to see business moves instead of chess ones.

Back then, our big idea for human and machine collaboration was what the chess nerds called a “Centaur.” You’d take one grandmaster, a person with all the intuition and strategic flair that comes from being a lumpy, biological creature, and you’d pair them with a top-tier chess engine. The human would handle the grand strategy, the unknowable art of the game. The machine would handle the brute force calculation, checking every move for blunders.

Together, they were unbeatable. Better than any human alone, and better than any computer alone. For a decade, this was the model. The human as the wise rider, the AI as the tireless horse.

It’s a lovely image. It is also, I must warn you, about as current as a steam ship.

From my perch as a fellow who invests in this sort of thing, I get to see the future arriving in bits and pieces, often in messy beta tests and bewildering technical papers. And the future isn’t a Centaur. It’s something vastly more complex, more powerful, and frankly, more interesting.

We are living at the dawn of the Centaur Swarm. And its making me thing an awful lot about the nature of intelligence itself.

I’ve captured my thoughts on this topic before:

But my thoughts have evolved after seeing the most recent technology.

To understand what that means, you first have to appreciate that we’ve been thinking about intelligence all wrong. We thought of it as one thing, a single scale of pure processing power. More horsepower in the brain meant more intelligence. But intelligence isn’t a single engine. It’s more like a chaotic, bustling kitchen full of specialists.

There’s a pattern recognition chef, a short-order cook for abstract reasoning, a sous chef for adapting to new goals, and a head waiter who handles all the social and emotional intelligence. A truly smart mind isn’t just one brilliant cook. It’s a perfectly coordinated kitchen staff.

Since 1956 and perhaps even before then we have tried to build one single AI that could do it all. The master chef. The results were impressive but always a little… lopsided. An AI could write a perfect sonnet but couldn’t figure out that a cat is not a suitable hat. It could identify a tumor in a scan with superhuman accuracy but couldn’t tell you if the patient was feeling sad about it.

So, the really clever people stopped trying to build a single master chef. Instead, they started building the whole kitchen staff.

This brings us to the new workforce.

At the heart of it is a creature called an AI agent. Now, “agent” is one of those Silicon Valley words that can mean anything and nothing. So let’s be simple. An AI agent is a tiny, self-contained specialist. It is a computer program with one very specific job, a goal it pursues with the relentless, unblinking focus of a hummingbird.

Imagine you could hire an employee who needed no sleep, no pay, no coffee breaks, and whose entire existence was dedicated to one task. Say, finding the absolute cheapest flight to Honolulu for the third week of June, considering weather patterns, airline reliability, and your known preference for an aisle seat. That’s an agent. You give it a goal. It goes off and does it.

You probably need to arm it with tools. There are a growing bank of tools that have been open-sourced. Every tool you add to this agent’s toolbox will increase its capabilities.

It’s a useful little thing on its own. Perhaps you can power it up with a few additional tools… but its real power comes when you stop hiring just one.

You hire ten. Or ten thousand.

This is the swarm. It’s a multi-agent system where all these little specialists can talk to each other, delegate tasks, and collaborate to achieve a goal far too complex for any single one of them. And this is where the Centaur model gets its spectacular upgrade.

The new system isn’t a rider and a horse.

It’s a three-tiered command structure, and it looks something like this: