OpenAI backs Isara’s agent-swarm bet

OpenAI has backed Isara, a nine-month-old San Francisco startup that raised $94 million at a $650 million valuation. That is a lot of money for a company with no product in market, but the pitch is even stranger: software that coordinates thousands of AI agents on hard analytical work.

The timing matters. In September 2025, Cognition, the company behind Devin, hit a $10.2 billion valuation with $73 million in annual recurring revenue. Investors are clearly willing to pay up for AI systems that do more than answer prompts. They want systems that can plan, split work, and finish tasks with less hand-holding.

Isara’s co-founders are both 23. Eddie Zhang is a former OpenAI safety researcher, and Henry Gasztowtt was studying computer science at the University of Oxford. The company says its demo already involves about 2,000 agents working together to forecast the price of gold. That is a neat demo, but the real question is whether the system can survive contact with messy, high-stakes work.

What Isara is actually building

Most current AI products still center on one model responding to one prompt. Isara is trying to move past that. Its system is designed to let hundreds or thousands of specialized agents communicate, divide labor, compare outputs, and converge on a shared answer.

That sounds elegant until you think about the failure modes. One agent hallucinates. Another agent copies the mistake. A third agent optimizes for the wrong objective. At small scale, this is annoying. At large scale, it can turn a forecasting system into a very expensive confidence machine.

The company’s early focus is investment firms, which would pay for predictive modeling software. Biotechnology and geopolitical analysis are secondary targets. Those are all markets where a better forecast can be worth a lot of money, but they are also markets where bad predictions can be embarrassing or costly very quickly.

• Funding raised: $94 million
• Valuation: $650 million
• Founding age: 9 months
• Current product: none in market
• Demo scale: about 2,000 agents
• Early target customers: investment firms

Isara is also hiring aggressively. The company has reportedly brought in about a dozen researchers from Google, Meta, and OpenAI. That matters because multi-agent systems are not just a product bet. They are an engineering and research bet, and the people building them need to understand where coordination breaks down.

For readers who have followed the rise of agent tooling, this is the next logical escalation. LangChain, CrewAI, and AutoGen already let teams coordinate a handful of agents on structured tasks. Isara is trying to stretch that idea into something far larger and less predictable.

Why OpenAI would write this check

OpenAI’s investment is interesting because it is backing a startup founded by one of its former researchers in a field adjacent to its own work. That is not a casual move. It is a bet that multi-agent coordination may become a core capability, and OpenAI wants visibility into the best ideas before they harden into a rival product line.

There is also a talent angle here. The AI industry has a habit of turning research teams into startups. Once the people with the strongest instincts leave, the parent company can lose more than headcount. It can lose a chunk of its future roadmap. Funding the spinout is cheaper than trying to win every talent battle.

“The AI industry’s most valuable resource is not compute but the researchers who know how to use it.”

That line appears in the reporting around this deal, and it captures the logic pretty well. OpenAI is not buying revenue here. It is buying optionality, relationships, and a chance to stay close to a technical direction it may care about later.

The size of the check also looks modest when you compare it with OpenAI’s own scale. The company is reportedly raising capital at a $300 billion valuation. A $94 million investment in a research-heavy startup is pocket change relative to that, especially if the technology becomes important.

There is precedent for this kind of behavior across the sector. Google, Microsoft, and Amazon have all used investments to keep close tabs on promising AI teams and the people behind them. The check is part financial exposure, part early warning system.

The money is big, but the technical risk is bigger

Isara is part of what The Information has called the “neolabs” wave: research-heavy AI startups founded by alumni of OpenAI, DeepMind, Anthropic, and Google Brain. These companies raise huge rounds before they have meaningful revenue because investors are betting on research output, not current sales.

That model can work, but it also creates a gap between demo and deployment that is easy to ignore when a pitch deck is strong. A demo with 2,000 agents forecasting gold is impressive. A production system that an investment firm trusts with real capital allocation is a different problem entirely.

The market is already crowded with AI platforms shipping agent features. The agentic AI market is projected to grow from $7.8 billion in 2025 to $52.6 billion by 2030. That is a large enough market to justify a lot of experimentation, but it also means incumbents have a path to close the gap if the underlying coordination problem turns out to be incremental instead of structural.

• Agentic AI market in 2025: $7.8 billion
• Projected market by 2030: $52.6 billion
• Cognition valuation in 2025: $10.2 billion
• Cognition ARR in 2025: $73 million
• OpenAI valuation: about $300 billion
• Isara round size: $94 million

That comparison is the real story. Cognition already has revenue. Isara has research talent, a demo, and a very large valuation. Investors are clearly willing to fund frontier bets before they become products, but the bar for proving utility will rise fast once the market asks for repeatable results instead of impressive experiments.

And this is where the multi-agent idea gets tricky. Coordination is hard even when the task is narrow. Once you ask thousands of agents to work on open-ended analysis, you have to solve for communication overhead, conflicting goals, error propagation, and evaluation. That is a lot to ask from a system that is still being defined.

What to watch next

Isara’s next move will tell us whether this is a serious company or just a very well-funded research project. If it can turn its agent-swarm demo into something investment firms actually pay for, then this round will look smart in hindsight. If not, the valuation will look like another example of capital chasing a compelling story.

The bigger question is whether multi-agent coordination becomes a category that startups own, or whether OpenAI, Anthropic, Google Gemini, and Microsoft Copilot absorb it into their own products before smaller labs can build a moat.

My read: the first real proof point will not be a flashy demo. It will be a customer who pays for forecasts, keeps paying after three months, and can explain why the system beats a human team on cost or speed. Until then, Isara is a high-upside research bet with a valuation that says more about investor appetite than product maturity.

If the company wants to justify the number, it needs to answer one simple question in public: what can 2,000 agents do that a smaller, cheaper system cannot?