Sakana AI launched Aquarium on Friday, giving enterprise agents a shared map of company data and rules. It packages work Sakana did by hand for financial clients. To scale as software, it must work without those engineers on site.
Sakana AI launched Aquarium on Friday to package the data and rules that enterprise agents need before they can work. The Japanese-language release names the ontology as the core of the platform. An ontology is a shared map of an organisation's data, rules and meaning.
Agreed definitions are a constraint on enterprise agents, and Sakana sells those first. Its approach resembles Palantir's, with engineers embedded in each client. The economics depend on reusing that work across clients without those engineers.
How Aquarium works
Aquarium rests on an ontology that links company data where it already sits. The Sakana release describes it as a structured map of knowledge, rules and data and how they relate. It links tables and documents without moving them from where they are stored. When a connected source changes, the change flows in, so agents decide on the updated picture.
Governance controls what an agent may do and keeps the record. It sets access permissions and approval steps before an action runs. It records everything from the grounds of a decision to its result. It also tracks speed, accuracy and model usage.
Managed agents then run tasks on what the ontology holds. The release says a workflow can combine looking up information, an AI judgement, running a function and a human approval. Describing it is enough to deploy it, and the platform manages the agent through its whole life.
Aquarium also has a fourth layer for data, according to its product page. It says a process is split into decision units and written as a workflow in YAML, a plain text configuration format. A human approval can sit inside the workflow, so a person signs off at a defined point.
The client work behind the product
The product comes out of Sakana's work for financial clients. The release lists an AI loan expert with MUFG Bank, proposal generation with SMBC Group and joint development with Daiwa Securities Group. That is three financial sector projects Sakana has already done.
In each, Sakana's forward deployed engineers learned the client's data and procedures. They handled workflow design, evaluation and accuracy improvement. Sakana says the data connection and terminology work was repeated from scratch in every project.
Aquarium packages that work into a shared base. Data and ontology built once can be used as they are by other agents. A client can start an agent for each operation quickly and manage all of them in one place. The name comes from many fish living in one tank.
The team's experience fits that approach. The team pairs Sakana's applied research engineers and forward deployed engineers with veterans of Palantir, Snowflake, Microsoft and AWS. Those veterans built data platforms and governance there. The approach is established; Sakana is applying it in a different language and market. Palantir's Foundry is the obvious comparison, though Sakana does not make it.
Approval and logging answer a buyer who must show why a decision was taken and who approved it. Sakana signed a defence ministry contract on 29 July, announced on 24 August. It covers a survey and demonstration of AI functions for integrated analysis.
A defence buyer would ask for that log, yet the contract does not mention Aquarium. Aquarium builds the log into the product instead of leaving it to each project.
Sakana's release pitches four uses for agents on the platform. Agents read documents and data to inform decisions. They track changes without staff collecting information again. They act under set permissions with a human approving where needed. And the record lets teams trace a decision and revise the procedure. Nikkei, reporting the launch on 23 October, says the platform records the basis for each AI decision.
Fugu orchestrates models, while Aquarium governs the agents that run on a customer's own data. Sakana's English page on Fugu Max and Ultra v2 says Fugu "has evolved from a beta thesis into an enterprise-grade orchestration engine".
Competition and unanswered questions
Aquarium faces established competitors. Hyperscalers and Palantir already sell ontology tools and agent governance. Sakana's launch has no price, no named customer and no availability date. Consultations are open, and case studies will follow in sequence.
The alliance with SCSK and Sumitomo Corporation, announced on 10 September, gives Sakana a Japanese integrator and a large partner network. SCSK is to bring integration, data preparation and security. Sumitomo brings about 900 consolidated companies and a customer base of 100,000.
The alliance's four initial areas do not name Aquarium. They are mission critical work in finance and manufacturing, cybersecurity, joint enterprise AI products and social infrastructure. The cybersecurity work uses Sakana's orchestration model. Neither release says Aquarium will be sold through the alliance.
The product page explains why Aquarium may still need substantial services work. It says the way Aquarium connects and syncs to a client's existing data platform is designed to suit each target. That is custom work.
Can it grow without more engineers?
The economics depend on how much work customers can reuse. The release says a second agent in the same organisation can reuse the ontology and connections. Then the extra agent costs only a workflow definition, and revenue follows licences.
If each new organisation needs its own engineers, as the financial projects above did, revenue follows headcount. Aquarium is then a consultancy with a good toolkit. Sakana has published no price or customer count, so the split cannot be measured today.
Sakana says corrections made in operation will feed back into the ontology, and that AI will help build agents. Both would cut engineer hours per project if they work. Sakana has given no date for either, and both are listed as work under development, not shipped features.
A published case study showing a second workflow on an existing ontology with little new Sakana staffing would support the platform model. A case study requiring a larger team on site would point to a services business. Sakana has set no date for case studies, so that test stays open past 23 October. The release says only that examples will be introduced in sequence. The first should help answer that question.
