OpenAI backs Isara, a nine-month-old startup with no product valued at $650 million

OpenAI has invested in Isara, a San Francisco startup that wants to teach thousands of AI agents to work as one analytical team. The company is only nine months old and has no product on the market yet, but the new round already values it at $650 million.

What Isara is building

Isara is not building yet another chatbot or a single "smart agent" for a narrow task. The company's idea is to run hundreds or thousands of specialized AI agents that split a large analytical job into parts, synchronize with one another, check goals against each other and assemble a shared result. The founders describe this as a shift from standalone tools to coordinated teams.

For the market this is an important change: the question is no longer whether a model answers a query, but whether a network of models can carry out a full research project. So far the most visible example is a demonstration in which about 2,000 agents jointly forecast the price of gold. It is easy to see from this where Isara wants to go commercially: expensive analytics, where even a small improvement in forecast quality can be monetized.

The first target segment is investment firms. After that, the startup is looking at biotech and geopolitics, and in the long term at systems that track global shifts and predict economic trends.

• About 2,000 agents have already been shown in a demonstration
• The first commercial focus is investment firms
• Next markets include biotech and geopolitical analysis
• The long-term goal is forecasting economic and political shifts

Why the round drew attention

The round's size, $94 million, already looks large for a company without a product. But the list of participants matters even more. OpenAI joined the deal, which immediately raises the stakes: the market reads the move as a signal that multi-agent systems are no longer a purely academic topic. The investors also named include Amity Ventures, Michael Eisenberg-style backers aside, Michael Ovitz, one of Uber's early backers, and billionaire Stanley Druckenmiller. Amid the general excitement around agentic AI, this turns Isara from a little-known team into a notable point on the industry map.

The founders add to the interest. Two 23-year-old co-founders launched the company in June 2025: former OpenAI AI-safety researcher Eddie Zhang and Oxford student Henry. Before launching, they co-wrote a paper for ICML 2024 on how AI systems can cooperate to improve decision-making in policy. That paper became the intellectual basis of the future product. Since then, Isara has hired about a dozen more researchers from Google, Meta* and OpenAI itself.

Where the main risk lies

Isara's biggest problem is not impressing investors but proving that its architecture works in a real environment. Even a single AI agent often makes mistakes on a complex task, loses context or behaves unstably. When there are hundreds or thousands of agents, new classes of failure appear: cascading errors, conflicting goals, accumulating hallucinations and communication noise. Existing frameworks such as LangChain, CrewAI and AutoGen can already coordinate small groups of agents, but that is a very different level of complexity from the open-ended analytical tasks at Isara's scale.

There is also a business gap between a striking demonstration and a product people will actually trust with money. Forecasting the gold price with a swarm of agents is a good showcase, but investment firms do not buy elegant architectures; they buy reliability, repeatability and clear accountability for results. Meanwhile the market is quickly filling with large players: Anthropic, Google and Microsoft are already adding multi-agent features to their platforms. So the next 18 months will be a test for Isara: either the startup shows that coordinating thousands of agents produces a qualitatively new result, or the idea stays an expensive research experiment.

What this means

Isara's story shows that investors are willing to value not only finished AI products very highly, but also strong research teams with an unconventional architectural bet. If the "agent swarm" model works in applied analytics, the enterprise AI market could gain a new class of tools. If not, it will be a telling example of how quickly capital inflates expectations around agentic systems.

*Meta is designated an extremist organization and banned in Russia.