Two 23-year-old researchers just convinced OpenAI to bet $650 million on an idea that sounds like science fiction: thousands of AI agents working together like a digital hive mind.
Isara, a San Francisco startup founded barely nine months ago, has pulled off what might be the fastest unicorn trajectory we’ve seen this decade. Eddie Zhang, a Harvard PhD candidate, and Henry Gasztowtt, an Oxford computer science student, published a paper in 2024 about AI systems cooperating to improve policymaking. Now they’re turning that academic curiosity into one of the hottest startups in Silicon Valley.
The Swarm Thesis
Here’s what makes Isara different from the AI agents everyone else is building.
When OpenAI releases Operator or Anthropic demos Claude controlling your computer, you’re watching a single agent—one AI tackling tasks sequentially. It’s impressive, but fundamentally limited. One model, one thread of execution, one bottleneck.
Isara’s bet is that the future isn’t a smarter solo agent. It’s thousands of dumber agents that communicate, coordinate, and emerge with collective intelligence that surpasses any individual model.
Think about how prediction markets work. No single person knows what Bitcoin will cost next month, but aggregate thousands of independent guesses weighted by conviction, and you get surprisingly accurate forecasts. Isara wants to do the same thing with AI.
Their early demos show this in action: thousands of agents coordinating to forecast gold prices. Each agent processes different information—macroeconomic indicators, geopolitical tensions, historical patterns, market sentiment. They communicate. They disagree. They reach consensus. And apparently, the results are impressive enough that OpenAI wrote a $94 million check at a $650 million valuation.
Why OpenAI Cares
Let’s be honest: OpenAI doesn’t invest in startups for portfolio diversification. They’re playing chess while everyone else plays checkers.
Three strategic angles here:
First, hedging the architecture bet. OpenAI built its empire on the “bigger model = better” thesis. GPT-4, GPT-5, whatever comes next—scale has been the moat. But what if that’s wrong? What if the next breakthrough comes from coordination, not scale? By backing Isara, OpenAI gets optionality. If swarms win, they have a stake in the winner.
Second, talent acquisition pipeline. Isara has been poaching researchers from Google, Meta, and—awkwardly—OpenAI itself. Rather than fight that brain drain, OpenAI turned it into an investment thesis. Now those researchers are building technology that could feed back into OpenAI’s ecosystem.
Third, platform play. If Isara’s agents run on GPT infrastructure, every swarm deployment becomes API revenue. OpenAI isn’t just betting on multi-agent futures; they’re positioning to be the substrate those agents run on.
The Technical Skepticism
I’ll be honest: I’m not fully sold yet.
The demo of agents forecasting gold prices is compelling theater, but it raises questions. How do you prevent agents from converging on groupthink instead of genuine wisdom of crowds? How do you handle adversarial agents—either bugs or intentional attacks—poisoning the swarm’s decisions? How do you explain the reasoning when thousands of agents contributed to a conclusion?
These aren’t new problems. They’re the same challenges that plague prediction markets, ensemble models, and distributed systems everywhere. Isara claims their coordination protocols solve them. The $650 million question is whether that’s true.
There’s also the compute economics. Running thousands of agents sounds expensive. Even with smaller models, coordination overhead adds up. Does the accuracy improvement justify the cost? For gold price forecasting, maybe. For everyday business applications, the unit economics need to work.
The Bigger Picture
What’s fascinating about Isara isn’t just the technology—it’s the timing.
We’re at a moment where the entire industry is questioning single-model supremacy. DeepSeek shocked everyone by training competitive models at a fraction of OpenAI’s cost. Reasoning models like o1 showed that longer inference time can substitute for larger training runs. Now Isara suggests that collaboration might substitute for capability.
Each of these represents a different escape route from the “scale is all you need” orthodoxy. And OpenAI is smart enough to realize that the winning approach might not come from their own labs.
The two founders are also worth watching. Eddie Zhang and Henry Gasztowtt are barely old enough to rent cars in most states, yet they’re recruiting researchers from trillion-dollar companies. That’s either the ultimate sign of a bubble—when any CS student with a paper can raise $100 million—or a sign that the technical insight is genuinely breakthrough-level.
What This Means For You
If you’re building with AI, here’s the practical takeaway:
Multi-agent architectures are moving from research curiosity to funded reality. The tools and frameworks for building agent swarms are about to get much better. If you’re working on problems that involve aggregating multiple perspectives—market analysis, risk assessment, strategic planning—keep an eye on this space.
If you’re an investor, the Isara deal signals that the “model layer” investment thesis is expanding. It’s no longer just about who builds the biggest model. The coordination layer, the orchestration infrastructure, the protocols that let agents work together—these are becoming investable categories.
And if you’re just watching from the sidelines, enjoy the show. We’re witnessing a fundamental debate about how artificial intelligence should be structured. Single powerful minds, or networks of simpler ones? The next few years will tell us who’s right.
Isara hasn’t shared much about their technology publicly yet. That’s probably intentional. When you’re sitting on a $650 million valuation and OpenAI’s backing before your first birthday, you don’t need press coverage. You need execution.
The founders went from academic paper to nearly-unicorn status in under a year. Now they have to prove that swarms of AI agents can do more than predict gold prices—they need to build something that changes how work gets done. That’s a much harder problem than raising money.
But if they pull it off? The single-agent paradigm everyone else is chasing might already be obsolete.