Sakana Aquarium: An Enterprise AI Agent Platform

What is Sakana Aquarium?

Sakana Aquarium is an enterprise platform from Sakana AI for building, running and operating AI agents. An AI agent is an AI system that, given a goal, plans the steps itself and uses tools to carry them out.

Aquarium gives those agents a shared, structured view of a company's knowledge and rules, then lets them check information, make a judgment and act. It is a platform, not a new model.

Sakana AI announced it on October 23, 2026

Sakana AI, the Tokyo AI lab, published the announcement on October 23, 2026. Both the blog post and the platform page are in Japanese only, so English readers have so far relied on secondary coverage. The company's earlier releases include Fugu, a system that combines several AI models to answer one request (related: Sakana Fugu Ultra is not always the best Fugu).

The announcement starts from one premise: every organization makes decisions on its own grounds and through its own steps. Aquarium is meant to turn that organization-specific context into a form an AI can work with.

The name refers to many fish in one tank

Sakana AI named the platform after the image of many fish swimming together in one well-kept tank. The fish are the agents. Other agents can reuse the data connections and ontology a company has already set up, so it can launch one agent per task and manage them all on the same platform.

Aquarium runs on three layers

The three layers of Sakana Aquarium

Layer	Role	Main functions
Ontology	Defines what the data means	Organizes knowledge, rules and data with their meanings and relationships, and links existing sources in place
Governance	Controls what agents may do	Sets access rights and approvals before execution, and records everything from the reasoning to the result
Managed Agents	Runs the work	Deploys agents from workflows and manages them from creation through operation

Aquarium is built from three layers: the ontology prepares the facts, Managed Agents carries out the work, and governance sets the limits. "The three layers of Sakana Aquarium" summarizes the role of each. The platform page also shows a fourth layer, Data, underneath them, which brings internal and external data and know-how together in a usable form.

The ontology links data without moving it

In Aquarium, the ontology is a structured model of the knowledge, rules and data behind a company's decisions, including what each item means and how the items relate. Agents use it to make judgments and run simulations.

Tables, Office documents and existing databases stay in the systems where they are stored. Aquarium links them to the ontology, and Sakana AI designs the connection and sync method for each customer's data platform. When a connected source changes, the ontology picks up the change, and agents act on the updated state instead of an old copy.

Governance sets permissions, approvals and audit logs

Governance is the layer that controls the data, the ontology and the agents. It covers four things:

• Access rights that decide who can use which data
• Approvals a person must give before an agent executes
• A record of each run, from the basis for a judgment to the final result
• Monitoring of execution speed, accuracy and model usage

Because every step is recorded, a team can trace what an agent looked at, what it decided, who approved it and what happened. Sakana AI expects companies to use those records to revise their procedures.

Managed Agents deploys workflows written in YAML

Managed Agents is the layer where agents carry out tasks using the information in the ontology. A team writes a workflow from four kinds of steps, and that workflow alone is enough to deploy the agent into live use:

1. Look up information
2. Let the AI make a judgment
3. Call a function, a prepared piece of code that performs an action
4. Ask a person for approval

The platform page shows this under the heading "Workflow as YAML." YAML is a plain-text format often used for configuration files.

In the example, a supplier's reply comes in, a load_reply step reads it, an analyze step has the agent judge it, and an approve step waits for a person before execution. Managed Agents also manages each agent's lifecycle, from creation through operation.

How Aquarium relates to Palantir's ontology

How Sakana AI and Palantir each describe their ontology

Point	Sakana Aquarium	Palantir
What the ontology models	Knowledge, rules and data behind decisions, with meanings and relationships	An enterprise's decisions, not just its data
Parts named alongside it	Governance and Managed Agents	Data, logic, action and security
Who wrote this	Sakana AI's announcement	Palantir's documentation

Aquarium uses the same central idea as Palantir: an ontology that models how a business makes decisions, which agents then work from. "How Sakana AI and Palantir each describe their ontology" puts each company's own description side by side. Sakana AI itself does not compare Aquarium with Palantir or any other product.

Both model decisions, not only data

Palantir calls its Ontology the system at the heart of its architecture and says it represents an enterprise's decisions, not simply the data. Sakana AI also describes Aquarium's ontology in terms of decisions. It holds the knowledge, rules and data that support decisions, and agents make their judgments from it.

The pieces around the ontology are named differently. Palantir describes decisions as an integration of data, logic, action and security. Aquarium splits the same ground into an execution layer (Managed Agents) and a control layer (governance).

Former Palantir and Snowflake engineers help build it

Some of the people building Aquarium previously worked on data platforms and data governance at Palantir, Snowflake, Microsoft, AWS and other companies. They work alongside Sakana AI's own engineers who have handled AI deployments for clients. Sakana AI has not said how Aquarium differs from Palantir's products in practice.

Why Sakana AI built Aquarium after its finance projects

Finance projects Sakana AI cites as earlier deployment work

Partner	Project
MUFG Bank	AI lending expert
SMBC Group	Automated proposal drafting
Daiwa Securities Group	Joint development

Aquarium grew out of Sakana AI's work putting AI into financial institutions and other fields that demand accuracy and accountability. Since its founding, Sakana AI has researched ways to make AI more capable without simply making models bigger: inference-time scaling, cooperation between several models and self-improvement. Alongside that research, it took on client projects, including the three in "Finance projects Sakana AI cites as earlier deployment work".

Every project repeated the same setup work

In every one of those projects, Sakana AI's team had to connect the data and organize terms and decision criteria again from scratch. That work fell to forward deployed engineers (FDEs), engineers who go on-site with a client to understand its data and procedures. They designed the workflows and handled evaluation and accuracy tuning.

Aquarium packages that repeated setup into a platform the whole organization can share. The design choice that lets one ontology serve many agents is the direct answer to this problem.

MUFG, SMBC and Daiwa are not Aquarium customers

The three finance projects appear in the announcement as Sakana AI's past implementation work, not as Aquarium deployments. A quick read of the announcement could suggest these banks already run Aquarium. The announcement does not say so, and Sakana AI says case studies will be introduced later.

What Sakana AI has not disclosed yet

What Sakana AI has disclosed about access (as of October 29, 2026)

Question	Official answer
Price	Not published
General availability date	Not published
AI models used	Not published
Named Aquarium customers	None yet
How to start	Consultation through the platform page

As of October 29, 2026, Sakana AI has not published Aquarium's price, a general availability date, the AI models it runs on or any named customer. "What Sakana AI has disclosed about access (as of October 29, 2026)" lists what the two official pages do and do not say.

No pricing or release date has been announced

Neither the announcement nor the platform page gives a price or a date when anyone can sign up. The only route the pages offer is a consultation request. That points to a sales-led product, with each deployment scoped together with the customer, rather than self-service.

No word on which AI models it uses

Sakana AI does not say which AI models power Aquarium's agents. The governance layer lets customers monitor model usage, but neither page names a model. They also do not say whether the agents run on Sakana AI's own models or on other vendors' models.

How to get started with Sakana Aquarium

Companies start by sending a consultation request through the contact form on the Sakana Aquarium platform page. Sakana AI says it shapes each consultation around the problem a company wants to solve and its operating conditions. From there, Sakana AI's FDEs build and test the setup together with the customer.

Start with a consultation and one workflow

The rollout begins with one operation and expands across the organization. That matches how the platform is built. The first workflow pays for the data connections and the ontology, and every later agent reuses them.

Two features are still in development

Sakana AI is developing two more capabilities. One takes in corrections and decisions made during daily operation and keeps the ontology and the agents up to date as the organization changes.

The other uses AI to help build the agents themselves. No release timing has been given for either.

Summary: Sakana Aquarium makes agent setup reusable

Sakana Aquarium is aimed at the preparation that comes before each agent, not at making agents smarter. In its finance projects, Sakana AI connected the data and defined business terms again for every client. Aquarium stores that work in one ontology, governed by one set of rules.

The first workflow pays for the connections and the definitions. Every agent added after it starts from that shared base instead of from zero.

Pricing, availability and the underlying models are still unknown. A team weighing Aquarium can prepare two answers before a consultation: which single workflow to start with, and at which step a person must approve. Those answers are useful on any agent platform, not only this one.