Sakana AI Releases Aquarium, an Ontology-Based Platform for Controlled AI Agents
Sakana AI has unveiled Aquarium, a platform where AI agents rely on a company's "dictionary": what its terms mean, how tables connect to documents, and which rules govern decisions. The ontology layer is built once and used by all agents. An agent checks the data, makes a decision, and carries out the action, pausing for human approval at the critical points. As a result, the new process does not require reconnecting data or reconciling terms from scratch.
AI-processed from Sakana AI Blog; edited by Hamidun News
Most enterprise AI agent projects get stuck not on the model but on the routine work around it. Data has to be connected, teams have to agree on what "active customer" or "overdue payment" means, permissions have to be defined for who can do what, and all of it has to be repeated for the next process. Sakana AI has set out to turn that routine into a product and has introduced Sakana Aquarium, a platform for building, deploying, and maintaining agents.
What's inside Aquarium
The foundation of the platform is an ontology. Put simply, it is a shared dictionary and map of the company: what kinds of entities exist (customers, contracts, orders), how they relate to each other, and what rules govern decisions about them. Tables, documents, and other formats connect to this map from existing storage, without data being migrated. When information is updated, the agent sees a fresh picture.
The platform has three parts:
- Ontology stores knowledge, rules, and data together with their meaning and the relationships between them.
- Governance handles control. It sets access permissions and mandatory approval before an action is taken. The system records what the agent relied on and what it ultimately did, and it shows the speed, accuracy, and model spending of its work.
- Managed Agents is the environment where agents run. A developer describes a workflow: reading data, a model decision, a function call, human confirmation. After that, the agent can be deployed, and the platform manages its lifecycle.
The name is explained with a metaphor: many fish swim in one aquarium with a well-maintained environment.
Why it matters in practice
The core idea is reuse. Connected data and an ontology built for one agent are also usable by others. Suppose a bank builds an agent to check loan applications. The next agent, say one that prepares offers for customers, does not have to be re-taught what a "customer" is or where to get their data. All agents run on one foundation and are managed from one place.
For product teams, the control scheme is also important. An agent acts within the permissions it has been granted, and at critical points a human confirms the step. Afterward, any decision can be reviewed through the log: which data was used, how the model reasoned, who approved it, and what the outcome was. In regulated industries such as finance, this kind of audit often matters more than polished demos.
Where this came from and what is still unknown
Sakana AI is known for research that improves AI quality without simply making models bigger: scaling at inference time, coordinated work among several models, and self-improvement. In parallel, the company has run applied projects in finance. These include an AI credit-screening expert with MUFG Bank, automated proposal generation with SMBC Group, and joint development with Daiwa Securities Group.
Those projects involved Forward Deployed Engineers, who worked through the customer's data and processes, designed workflows, evaluated results, and improved accuracy. The company noticed that connecting data and reconciling terms had to be done from scratch in every project. Aquarium packages that work into a shared platform. The product is being built by engineers from Sakana AI along with people from Palantir, Snowflake, Microsoft, and AWS, that is, from companies that build data platforms and data governance tools. The ontology-based approach resembles what Palantir does, but that is our comparison, not Sakana's own wording.
There are limitations too. The announcement includes no benchmarks, no pricing, no list of supported data sources, and no names of the models used inside the platform. There are also no public case studies with measurable results; Sakana says it will share them later. Based on the text, the product is sold through consulting and implementation involving FDEs, rather than as a service that can be picked up and tried independently. Looking ahead, the company names two directions: updating the ontology and agents based on corrections and decisions that emerge during operation, and using AI to help build the agents themselves.
The takeaway for developers and founders is straightforward. The market for agent platforms is shifting from "which model should we use" to "what does the agent base its decisions on" and "how do we control it." If you are building agents for several processes, it is worth planning a shared layer of definitions and access rights in advance, or you will pay for repeated setup in every project. Whether Aquarium actually helps in real-world work can only be judged once the first public cases appear.
- Nov 21, 2026Apex Compute Releases DLSS Alternative for GTX 16 and Radeon: Frame in 1.01 ms vs 1.26 ms
- Nov 21, 2026Mistral Opens Singapore Hub for Asia, Plans to Grow to 200 Staff by 2027
- Nov 21, 2026Anthropic Heads to IPO at $1.4 Trillion, but Seven Founders Keep 50.1% of the Votes
- Nov 21, 2026AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens
Key Questions Answered
What's inside Aquarium?
The foundation of the platform is an ontology. Put simply, it is a shared dictionary and map of the company: what kinds of entities exist (customers, contracts, orders), how they relate to each other, and what rules govern decisions about them. Tables, documents, and other formats connect to this map from existing storage, without data being migrated. When information is updated, the agent sees a fresh picture.
Why it matters in practice?
The core idea is reuse. Connected data and an ontology built for one agent are also usable by others. Suppose a bank builds an agent to check loan applications. The next agent, say one that prepares offers for customers, does not have to be re-taught what a "customer" is or where to get their data. All agents run on one foundation and are managed from one place. For product teams, the control scheme is also important.
Where this came from and what is still unknown?
Sakana AI is known for research that improves AI quality without simply making models bigger: scaling at inference time, coordinated work among several models, and self-improvement. In parallel, the company has run applied projects in finance. These include an AI credit-screening expert with MUFG Bank, automated proposal generation with SMBC Group, and joint development with Daiwa Securities Group.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.