By Hamza Faisal
Sakana AI is coming for one of the most expensive problems in enterprise AI: every new agent a company deploys has to relearn the business from scratch. On October 23, 2026, the Tokyo-based AI lab introduced Aquarium, a managed platform that packages business data connections, definitions, decision rules, and operating controls into reusable infrastructure that any agent inside a company can draw on.
The pitch is simple enough. Instead of wiring up each agent's data sources, permissions, and judgment criteria individually, a company builds that foundation once in Aquarium — and every subsequent agent inherits it. Sakana says adoption will begin through customer-specific consultation, with the company working jointly with early customers to build and validate the first deployment, typically around a single business operation. Neither pricing nor a general-availability date has been announced.
An ontology at the center
At the heart of Aquarium sits an ontology: an organized map of business knowledge that spells out what company terms mean, how data relates across systems, and which decision criteria apply. Data connections prepared for one agent can support others on the same foundation. The platform connects information in formats including tables and documents while leaving source data in its existing storage locations, with updates reflected for agents to use.
Workflows built on the platform can combine information retrieval, AI judgment, and function execution — with human approvals and records of decisions and results. Access permissions and approval requirements are set before execution, while records trace the information consulted, the decisions made, the approvals received, and the outcomes. Sakana also says Aquarium gives companies visibility into execution speed, accuracy, and model usage. AI-assisted agent construction and updates based on operational feedback remain planned capabilities rather than shipping features.
Finance projects as the foundation
Sakana says Aquarium's development draws on earlier financial-sector projects with Mitsubishi UFJ Bank, SMBC Group, and Daiwa Securities Group — work the company describes as groundwork for the platform rather than new Aquarium deployments. That lineage matters: banks are among the strictest test environments for AI, with tightly controlled data access and zero tolerance for agents acting on misunderstood definitions.
The announcement lands at a moment when the whole industry is racing to build shared infrastructure for agents. Anthropic donated its Agent Skills standard to the Linux Foundation's Agentic AI Foundation just this week, backed by OpenAI, Hugging Face, and Alibaba projects — while banks like JPMorgan are simultaneously tightening which agents get to touch customer accounts. Sakana, founded by former Google researchers David Ha and Llion Jones, has spent the year building an enterprise story on top of its research reputation: the fully automated AI Scientist, the MASS multi-agent self-improvement framework announced earlier this month, and now a platform meant to make agents safe, auditable, and — critically — not start from zero every time.
By Hamza Faisal
