$650M Says Agent Coordination Is the Problem. The Question Is How You Solve It.

Two 23-year-olds just raised $94 million to build AI agent swarms. OpenAI backed them at a $650 million valuation. The company is called Isara, it was founded nine months ago, and it has about fourteen people.

When OpenAI puts money behind a company whose entire thesis is "agents need to work together," it tells you where the industry thinks the next layer of value sits. The model race is becoming table stakes. The coordination layer, how agents divide work, share context, and produce outputs no single model could generate alone, is where things are heading.

What Isara is actually building

Isara is what's being called a "neolab", a research startup built on top of academic work. The founders, Eddie Zhang and Henry Gasztowtt, co-authored a paper at ICML 2024 exploring how AI systems could cooperate to improve policymaking. That paper is the intellectual foundation. The team now includes about twelve researchers hired from Google, Meta, and OpenAI.

Their core focus is building the architecture that allows different AI agents to align goals, exchange data, and divide tasks across complex processes. The founders describe the vision as agents functioning like a well-coordinated team rather than isolated tools.

The initial target is investment firms, predictive modelling software, with biotechnology and geopolitical analysis as secondary markets. The longer-term ambition is training agent swarms to collaborate on problems like tracking geopolitical shifts or forecasting economic trends.

It's a research-first bet. Build the theory of multi-agent coordination, then find the applications.

Two schools of thought

There's a tension emerging in this space between two approaches.

The first is what Isara represents: start from research, build general coordination protocols, and apply them to complex analytical domains. Big funding, academic pedigree, long time horizons.

The second starts from the other end: pick a specific production problem where multi-agent coordination is clearly useful, build the orchestration layer to solve it, and expand from there. Engineering workflows are a natural fit — tasks decompose cleanly, feedback loops are tight, outputs are verifiable. Platforms like Workforce AI have been running this architecture in production, with agents picking up tickets, writing code, reviewing PRs, and merging.

Both paths are legitimate. But they're solving for different timelines and different risk profiles.

Swarms vs hierarchies

Isara's language centres on "swarms" — the idea that enough agents with the right communication protocols will produce emergent intelligence. It's a compelling vision, and there's real research behind it.

But the production evidence so far tells a more nuanced story. A Google DeepMind paper found that agent teams often perform worse than a single agent working alone, unless the work decomposes into independent tasks and there's a clear hierarchy. Stanford research shows agents default to unhelpful consensus-seeking in flat structures. And Evan Ratliff's Shell Game experiment, where AI agents were tasked with running a tech company, showed them burning through API credits on small talk rather than getting work done.

None of this means swarm approaches can't work. It means the engineering constraints around multi-agent systems are real, and solving them at the protocol level is genuinely hard. Whether Isara's research-first approach cracks it faster than the iterate-in-production approach remains to be seen.

What the money tells us

Regardless of where you sit on the swarm-vs-hierarchy question, the Isara raise validates a few things worth noting.

Agent coordination is now a funded category. It's not a feature of a model company or a thin wrapper. Investors are pricing it as standalone infrastructure.

The market is bifurcating. Research labs targeting complex analytical problems on one side. Production-focused platforms targeting engineering and operational workflows on the other. Both will coexist. The question is which delivers measurable value first.

Model providers are paying attention. OpenAI backing Isara is strategic. If agent coordination becomes the dominant way enterprises use AI, the coordination layer accumulates leverage over the model layer. Every major provider will want exposure to this.

Isara is part of a broader wave — investors have reportedly put or discussed $2.5 billion across five "neolab" startups in just over a month. The appetite for foundational AI research outside the major labs is clearly there.

The question that matters

The exciting version of this story is swarm intelligence — thousands of agents collaborating on geopolitical analysis, drug discovery, and economic forecasting. That's the vision that raises $94 million.

The practical version is quieter. Can you get a team of agents to reliably divide a complex task, execute the pieces independently, and produce a verified output? That's the question the entire category needs to answer, whether you're building from research or from production.

It'll be worth watching which approach gets there first.