Sakana AI launches Aquarium: an ontology layer that lets organizational agents act, with approvals and audit trails

Sakana AI released Sakana Aquarium, a platform for building, running and operating organizational AI agents, aimed at turning scattered organizational information into usable premises for decisions. In earlier finance-related projects that demand accuracy and accountability, connecting data and organizing terminology and judgment criteria had to be repeated from scratch by forward-deployed engineers for each case. Sakana AI cites prior deployments including an AI lending expert with MUFG Bank, automated proposal generation with SMBC Group and joint development with Daiwa Securities Group; Aquarium consolidates that preparation work into shared organizational infrastructure.

The platform has three parts, ontology, governance and a managed-agent runtime, and agents perform information lookup, AI judgment, function execution and human approval on top of the ontology. The ontology systematizes knowledge, rules and data together with their meanings and mutual relations, and connects differently formatted information such as tables and documents without relocating it from existing stores, with updates reflected. Governance sets access permissions and pre-execution approvals, records the chain from judgment basis to execution result, and tracks execution speed, accuracy and model usage; the runtime lets teams describe lookup, judgment, function execution and human approval as a workflow and deploy it, managing the agent lifecycle.

Data connections and an ontology prepared once can be reused directly by other agents, so agents for individual operations can be launched quickly and managed on the same foundation. This targets the practice of building a separate agent for each operational step and repeatedly reorganizing data and judgment criteria. Sakana AI says development involved its ARE and FDE staff who handle AI deployment, together with members who built data foundations and data governance at Palantir, Snowflake, Microsoft and AWS.

The report covers Sakana AI's public description of Aquarium's mechanisms, namely the ontology, governance and managed-agent runtime and how they are reused; the material gives no deployment scale, performance figures or customer measurements.

Sakana AI says Aquarium will keep adding features and is developing mechanisms to fold corrections and judgments from operations back into ontology and agent updates, and to have AI assist agent construction; these remain directions under development, and the material does not state their completion status.