Sakana AI Launches Aquarium Agent Platform

Sakana AI announced Sakana Aquarium, a platform for building and operating AI agents that support organizational decision-making and execution. It structures each organization’s decision-making assumptions in a form AI can use, then runs agents that handle information checks, decisions, and actions. The platform is built around three components: Ontology, Governance, and Managed Agents.

Ontology organizes the knowledge, rules, and data behind decisions, including their meanings and relationships. It connects information in different formats, such as tables and documents, without moving it from existing storage, and reflects updates so agents can make decisions using current information. Governance controls data, the ontology, and agents by setting access permissions and approvals before actions. It records the basis for decisions through execution results, and tracks execution speed and accuracy as well as model usage.

Managed Agents provide the foundation for tasks based on information in the ontology. Teams can deploy agents by defining workflows that combine information retrieval, AI decisions, function execution, and human approval; the platform manages the agents’ full life cycle. Prepared data and ontologies can be reused by other agents, allowing teams to launch agents for different tasks and manage them on the same platform. Sakana AI named the platform Aquarium after the image of many fish swimming in a shared, prepared environment.

Sakana AI says the platform lets AI read an organization’s data and documents so it can draw on a wider range of information for decisions. It says reflecting updates helps reduce the effort of gathering information from different locations, leaving people to focus on decisions and responses. Under defined permissions and rules, AI can handle tasks from checking information through execution, with people approving steps where needed. Records let users trace the information consulted by AI, its decisions, approvals, and execution results, and use them to review procedures.

Since its founding, Sakana AI has researched ways to improve AI capabilities beyond simply making models larger, including inference-time scaling, coordination among multiple models, and self-improvement. It has also worked on implementations in fields such as finance, where accuracy and accountability are required. Its projects include an AI lending expert with Mitsubishi UFJ Bank, automated proposal generation with the SMBC Group, and joint development with the Daiwa Securities Group. In each project, FDEs—engineers who work on site to support implementation—learned the data and procedures, then handled workflow design, evaluation, and accuracy improvements.

Aquarium brings recurring work on data connections, terminology, and decision criteria into a shared platform for use across an organization. Its development team includes ARE and FDE staff who handled implementation at Sakana AI, along with members who built data platforms and governance systems at Palantir, Snowflake, Microsoft, and AWS. Sakana AI says it plans to develop ways to incorporate operational corrections and decisions, update ontologies and agents as organizations change, and use AI to assist with agent building. The company plans to have FDEs build and validate systems with customers, expand their use from individual tasks across organizations, and introduce use cases in stages.

Sakana AI announced Sakana Aquarium, a platform for building and operating AI agents that support organizational decision-making and execution. It structures each organization’s decision-making assumptions in a form AI can use, then runs agents that handle information checks, decisions, and actions. The platform is built around three components: Ontology, Governance, and Managed Agents.

Ontology organizes the knowledge, rules, and data behind decisions, including their meanings and relationships. It connects information in different formats, such as tables and documents, without moving it from existing storage, and reflects updates so agents can make decisions using current information. Governance controls data, the ontology, and agents by setting access permissions and approvals before actions. It records the basis for decisions through execution results, and tracks execution speed and accuracy as well as model usage.

Managed Agents provide the foundation for tasks based on information in the ontology. Teams can deploy agents by defining workflows that combine information retrieval, AI decisions, function execution, and human approval; the platform manages the agents’ full life cycle. Prepared data and ontologies can be reused by other agents, allowing teams to launch agents for different tasks and manage them on the same platform. Sakana AI named the platform Aquarium after the image of many fish swimming in a shared, prepared environment.

Sakana AI says the platform lets AI read an organization’s data and documents so it can draw on a wider range of information for decisions. It says reflecting updates helps reduce the effort of gathering information from different locations, leaving people to focus on decisions and responses. Under defined permissions and rules, AI can handle tasks from checking information through execution, with people approving steps where needed. Records let users trace the information consulted by AI, its decisions, approvals, and execution results, and use them to review procedures.

Since its founding, Sakana AI has researched ways to improve AI capabilities beyond simply making models larger, including inference-time scaling, coordination among multiple models, and self-improvement. It has also worked on implementations in fields such as finance, where accuracy and accountability are required. Its projects include an AI lending expert with Mitsubishi UFJ Bank, automated proposal generation with the SMBC Group, and joint development with the Daiwa Securities Group. In each project, FDEs—engineers who work on site to support implementation—learned the data and procedures, then handled workflow design, evaluation, and accuracy improvements.

Aquarium brings recurring work on data connections, terminology, and decision criteria into a shared platform for use across an organization. Its development team includes ARE and FDE staff who handled implementation at Sakana AI, along with members who built data platforms and governance systems at Palantir, Snowflake, Microsoft, and AWS. Sakana AI says it plans to develop ways to incorporate operational corrections and decisions, update ontologies and agents as organizations change, and use AI to assist with agent building. The company plans to have FDEs build and validate systems with customers, expand their use from individual tasks across organizations, and introduce use cases in stages.