Sakana AI launches Aquarium, a platform that runs company agents on an ontology

What Sakana AI announced

On 23 October 2026 the Tokyo AI company Sakana AI announced Sakana Aquarium, a platform for building, running and operating AI agents inside organisations. Sakana describes it as a system that supports both the decisions an organisation makes and the work of carrying them out. The announcement is written in Japanese, and the company has also put up a product page with a short tour of the features.

The core idea is simple to state. Every company decides things differently: what evidence it trusts, which rules apply and in what order steps happen. Sakana says Aquarium turns those organisation-specific assumptions into a form an AI can work with, and then runs agents on top of that shared base. The agents check information, make a judgement and carry out the task, with people approving where the company requires it.

The name is a metaphor that Sakana spells out itself. Sakana means fish in Japanese, and the company says it chose Aquarium because many fish can swim in one tank whose environment is kept in good order. In the same way, data and definitions prepared once are meant to be reused by many separate agents running on one platform.

Ontology, governance and managed agents

Sakana names three mechanisms at the centre of Aquarium. The first is the ontology. In this setting an ontology is a structured map of an organisation's knowledge, rules and data, with the meaning of each item and how the items relate to each other. According to Sakana, it connects information in different formats, such as tables and documents, without moving it out of the systems where it already lives. When the connected sources are updated, the ontology reflects the change, so agents make decisions on the latest picture.

The second is governance. Sakana says this layer controls the data, the ontology and the agents. Administrators set access permissions and approval steps that must happen before an action runs. The platform records everything from the grounds for a decision through to the result of the action, and it also tracks how fast and how accurately tasks run and how the underlying AI models are being used.

The third is managed agents, the runtime. Agents carry out tasks using what the ontology holds. Sakana's release says a team can describe a workflow that combines looking up information, an AI judgement, running a function and a human approval, and deploy it from that description. The platform then manages the agent's whole life cycle. The product page adds a fourth label, data, for connecting internal and external data and know-how in a usable form.

Workflows written as YAML, with a human step

The Aquarium product page gives the most concrete view of how work is defined. Under the heading Workflow as YAML, it says a business process is split into units of decision and written as a workflow an AI can execute, and that human approval can be built into that workflow. YAML is a plain-text configuration format that many engineering teams already use for deployment files.

The page illustrates this with a small example that handles a reply from a supplier. Two inputs, the supplier's reply and the supplier record, feed a step that loads the reply, then a step where an agent analyses it, then a human approval step before anything is executed, shown with a waiting-for-approval status. It is a simple diagram, but it shows the design choice: the approval is a named step inside the process, not an afterthought bolted onto a chat window.

Two other parts of the page describe the data side. Connect in place says tables, Office documents and existing databases are linked through common definitions so search, analysis and agents can use them, with the connection and sync method designed for each target. Stay updated says changes in the connected sources flow into the ontology so agents act on current information.

Built from bank projects

Sakana presents Aquarium as the product of its consulting-style work in finance, a field where accuracy and accountability matter. The release lists three examples: an AI lending expert with MUFG Bank, automatic generation of proposal documents with SMBC Group, and joint development with Daiwa Securities Group. Nikkei, reporting the launch the same evening, also pointed to the company's work with large financial institutions on proposal writing and loan support.

In each of those projects, Sakana says, its forward deployed engineers learned the customer's data and procedures and handled everything from workflow design to evaluation and accuracy improvements. Forward deployed engineers are staff who work on site with a client to build the system. The company says a pattern emerged: every project repeated the same groundwork of connecting data and sorting out terms and decision criteria from scratch. Aquarium packages that groundwork into a base the whole organisation can share.

Sakana says the platform is being built by its own AI engineers and forward deployed engineers together with staff who previously built data platforms and data governance at Palantir, Snowflake, Microsoft and AWS. That is a hint about the intended shape: closer to an enterprise data platform with agents on top than to a chat assistant.

How it fits the enterprise agent race

Sakana is better known for research. It has worked on scaling at inference time, getting several models to cooperate and self-improvement, and earlier this month it open-sourced The Agentic Scientist, a research agent built on coding agents. Aquarium is a different kind of release: a commercial platform aimed at everyday operations rather than a model or a paper.

The pitch puts Sakana in a crowded field. Microsoft is giving Copilot Studio agents their own Microsoft 365 accounts and shared memory, and Google Cloud has launched Agent Compute Engine, a microVM for every agent. The idea of mapping company data into an ontology before letting agents act on it also has precedents, including Helix Foundry's local ontology workspace. Sakana's own staffing note about former Palantir engineers points to the best-known enterprise ontology approach, although Sakana does not make that comparison itself.

The approval step is the part most buyers will look at first. Recent agent launches have converged on the same answer: the agent prepares the action and a person signs off, as with Stripe's agent payments, which need human approval by default. Our explainer on approval gates for AI agents covers why putting that check inside the workflow matters.

What Sakana has not said

A lot is still missing. The Japanese AI news site TRAIL went through the release and listed what it does not say: price and contract terms, a start date for service, which companies and regions can use it, which AI models it runs on, how it connects to other vendors' systems and data stores in detail, and any named customer deployment. AI Lookout's reading of the release and product page matches that list.

For now, adoption works by consultation. Sakana says it will take enquiries based on the problem a customer wants to solve and its operating conditions, and that its forward deployed engineers will build and test with customers, expanding from one operation to the wider organisation. It says case studies will be introduced in turn, without a date.

AI Lookout has not used Aquarium. The descriptions above come from Sakana AI's announcement and product page, read on 25 October 2026, with TRAIL and Nikkei as secondary sources. Claims about speed, accuracy and reuse are the company's descriptions of its design, not measured results.

What to watch next

Sakana says it will keep extending the platform. Two features are named as in development: a mechanism that takes in the corrections and decisions made during operation and keeps updating the ontology and the agents as the organisation changes, and a mechanism in which AI helps build the agents themselves.

The first real test will be the promised case studies. A platform like this lives or dies on how much of the setup work actually carries over from one agent to the next, which is exactly the claim Sakana makes and has not yet shown publicly. Companies weighing it should ask who owns the ontology, how audit records can be exported, and who signs off on each automated step, the same questions our guide to who owns AI governance recommends.