Sakana unveils Aquarium to connect enterprise rules with AI agent workflows

Sakana AI announced Aquarium on October 23, presenting a platform for building, running and maintaining agents around an organization’s own decision-making processes. The system organizes scattered information into shared business definitions before connecting it to agent actions. Its public pages direct enterprises to deployment consultations; a general-availability timetable and pricing have not been published.

The foundation is an ontology: a structured account of knowledge, rules and data, including their meanings and relationships. The product page describes linking tables and documents through common definitions while keeping existing data sources in place. Connection and synchronization methods are designed around the systems involved. Updates to connected information are reflected in the ontology, giving agents a revised context for judgments and simulations.

Aquarium’s managed execution layer puts that context into workflows. Business processes are divided into decision steps and expressed in YAML, with human approval available inside the workflow. The announcement describes information retrieval, AI judgments, function execution and approval as elements that can be combined for deployment. Managed Agents handles the agent lifecycle, while prepared data connections and definitions can be reused when another operation needs its own agent.

Governance covers the data, ontology and agents together. Organizations can set access permissions and approvals before execution, and trace the information consulted, decisions made and resulting actions through records. Visibility covers execution speed, accuracy and model usage. Records help teams review procedures and understand an agent’s action.

Deployment includes work by Sakana’s forward deployed engineers, or FDEs, alongside customers. The company’s description of that role emphasizes understanding existing systems, business constraints and data before designing a solution. Engineers develop prototypes, connect dispersed information and analyze problems that emerge in operation. Requirements and lessons from individual projects are fed back to the platform team, linking customer-specific implementation with changes to the shared product.

Earlier enterprise projects provide context for that approach. On April 30, 2026, Sakana announced a proposal-generation application with SMBC Group, following a partnership established in May 2025. The application was beginning practical use at Sumitomo Mitsui Banking Corporation. Agents with different roles coordinate research, analysis, hypothesis building, narrative development, quality assessment and fact-checking. The project combines financial and nonfinancial company information to support corporate banking proposals. These are earlier application-development efforts cited in Aquarium’s announcement.

A March 19, 2026 interview about Sakana’s MUFG lending assistant described another implementation process. The team worked across initial analysis, financial simulations and drafts of internal approval documents as the project moved toward evaluation on real cases. Staff reported nearly 1,500 pieces of feedback during evaluation. They described using AI to classify issues, extract improvement themes and revise prompts and rulebooks, while comparing automated evaluations with bank employees’ assessments. That work documented how institutional judgment was translated into application workflows.

Aquarium’s roadmap extends those ideas into platform capabilities still under development. Sakana is building mechanisms that incorporate operational corrections and decisions so the ontology and agents can keep changing with the organization. It is also developing AI assistance for constructing agents themselves. The company plans continued joint construction and validation with customers, expanding from individual operations, and says Aquarium use cases will be introduced progressively.