Sakana AI unveils Aquarium: one shared playbook for enterprise AI agents

• Who: Sakana AI, the Tokyo frontier-AI lab
• What: Aquarium — a managed foundation that lets enterprise AI agents share data connections, definitions, decision rules, and operating controls
• When: introduced October 23, 2026
• Core tech: a business ontology organizing terminology, data relationships, and decision criteria; connects tables and documents while source data stays in existing storage
• Governance: access permissions and approvals set before execution; records trace the information consulted, decisions made, approvals received, and results; metrics on execution speed, accuracy, and model usage
• Commercial reality: adoption via customer consultation and joint validation around one operation — no pricing, no general-availability date announced

Sakana AI introduced Aquarium today, a managed foundation for enterprise AI agents that lets companies reuse the business knowledge built for one agent across all of them. Data connections, definitions, decision rules, and operating controls prepared for one deployment can support others on the same foundation — the point, the company says, is to stop rebuilding the same business knowledge for every new agent.

The ontology is the product

At the center of Aquarium is an ontology: an organized map of what business terms mean, how data relates, and which decision criteria apply. Aquarium connects information held in tables and documents while leaving source data in its existing storage — and updates to that knowledge are then reflected for every agent to use.

That framing matters because it inverts the usual enterprise-agent pitch. Vendors tend to sell the agent; Sakana is selling the shared ground the agents stand on. If every agent consults the same definitions and the same decision criteria, the tenth agent a company deploys starts from the company's accumulated knowledge instead of from zero.

Approvals and audit trails are built in

Aquarium's workflows combine information retrieval, AI judgment, and function execution — with human approval gates. Access permissions and approval requirements are set before execution, and records trace the information consulted, the decisions made, the approvals received, and the results. The company also promises visibility into execution speed, accuracy, and model usage.

The emphasis is the tell. The bottleneck for enterprise agents is rarely the model — it is the governance around it: who may see what, which decisions need a human sign-off, and how to prove what the agent did afterward. Aquarium is Sakana packaging that layer as infrastructure instead of bespoke implementation work.

Finance experience, not finance deployments

The platform draws on Sakana's earlier finance-sector projects with MUFG Bank, SMBC Group, and Daiwa Securities Group. The company is careful to describe that work as groundwork for the platform rather than newly announced Aquarium deployments — none of these banks is a customer going live today.

Nor can anyone buy Aquarium off the shelf yet. Adoption starts with customer-specific consultation and joint construction and validation, initially scoped around a single operation. Sakana has disclosed neither pricing nor a general-availability date, and capabilities like AI-assisted agent construction and updates based on operational feedback are described as planned, not shipped.

Still, the direction of travel is clear. The agent story is moving from demos to durable operations, and whoever sells the operating layer — shared definitions, shared rules, shared auditability — owns the boring parts that determine whether agents survive contact with a real company. Aquarium is Sakana's bid for exactly that layer.