Sakana Aquarium: To Get the Enterprise Agent Up and Running, First Clarify the Organizational Rules

AI in the corporate world is evolving from answering questions to executing tasks. However, a simple request like “Help me process this order” may simultaneously involve customer credit, inventory criteria, pricing authority, and approval responsibilities. Just because a model can read this sentence doesn’t mean it understands all these conditions.

October 23, 2026, Sakana AI Release: Sakana Aquarium...brings organizational knowledge, ontologies, governance, and agent operations onto a single platform. The approach it proposes is worth noting: organizing the premises on which corporate decisions are based into reusable infrastructure, so that different agents do not have to guess anew each time “how this company operates.”

I. The real challenge is that the same word can have different meanings in different departments.

Let’s take a hypothetical order scenario as an example. A salesperson seeing “In Stock” might interpret this as meaning they can provide a quote; the warehouse’s inventory includes goods awaiting quality inspection; and the finance department is concerned about whether ownership of the goods is clear and whether the customer has any outstanding payments. If the Agent retrieves data from only one of these tables, the response may be syntactically correct and numerically accurate, but it could result in a commitment that is not feasible from a business perspective.

Problems like these cannot be left entirely to prompts to solve. Prompts can alert the model to credit risks, but they cannot decide on behalf of the company which table is the authoritative source for credit limits, nor can they automatically resolve discrepancies in the definitions of “inventory for sale” and “book inventory.” Companies need to clearly define entities, relationships, states, and rules before allowing the model to operate within this shared framework.

The ontology used by Aquarium can be understood as a structured expression of this common language. It does more than simply tag documents; it organizes knowledge, rules, data, and their relationships to serve as the foundation for decision-making and action. The significance of this design lies in the fact that when the next Agent takes over a different business process, they can reuse existing definitions rather than recreating a set of incompatible interpretations.

II. There is still a missing link in the chain of responsibility between “finding the information” and “completing the task.”

ground Aquarium Product OverviewThe platform connects to existing data, organizes information using an ontology, and expresses business processes through workflows; manual approvals can be embedded within these workflows. The specific methods for connecting and synchronizing data must be designed based on the target data environment. The release notes also outline governance capabilities such as permissions, pre-execution approvals, decision criteria, and result logging.

This has changed the way companies evaluate agents. In the past, demonstrations often focused on whether an agent could answer a question correctly; now, once the agent is integrated into the business system, at least the following questions must also be asked:

• Which version of the data is being read, and is the latency acceptable?
• Which rules were used in the derivation, and which were simply model-based judgments?
• Who can approve this action, and does the approval cover the current amount and scope?
• How can I recover after an operation fails? Can the log pinpoint the specific step where the problem occurred?

Suppose a procurement agent finds a cheaper supplier. A truly valuable workflow should not stop at a “successful recommendation,” nor should it place an order on behalf of the company directly. It needs to verify the supplier’s eligibility, delivery capabilities, and budget, and then pass on the decision requiring a manager’s approval. There should be a clear chain of responsibility between procurement recommendations and procurement commitments.

III. Shared Infrastructure: More Important Than Simply Adding More Agents

A company can quickly acquire multiple assistants for sales, customer service, finance, and other areas, but this does not necessarily lead to effective collaboration. If customer IDs cannot be matched, pricing rules are maintained in separate documents, and approval records are scattered across chat windows, the more agents there are, the higher the administrative burden may actually become.

Therefore, the commercial value of Aquarium is best understood in terms of reuse: Can the customer definitions, permission boundaries, and data connections that have been set up in one workflow be reused in the next workflow? When rules change, can they be updated centrally rather than by modifying each prompt individually? These questions go beyond simply completing a few more tasks than shown in the demo and are more closely related to long-term operational costs.

From an organizational perspective, this also means that projects cannot be driven solely by the AI team. The data team provides reliable sources, business leaders explain exceptions, the security team defines permissions, and operations staff provide examples of failures. The platform can host the rules, but the enterprise itself must own them.

IV. The cost most often underestimated is that of turning implicit experience into explicit criteria

Long-time employees know which customers should receive priority shipping and which exceptions must be escalated—knowledge that is often not documented in the system. When this information is formalized, companies may find that different managers give different answers to the same question.

This isn’t simply data cleansing; it’s the manifestation of a management issue. If business definitions haven’t been standardized yet, rushing to have the Agent execute tasks may simply automate the original vague judgments. Conversely, clarifying the rules may become the project’s most valuable output, even if only a small number of actions are automated in the first phase.

It is also important to distinguish between “updates being reflected” and “absolute real-time at all times.” The update frequency of the business source system, the handling of synchronization failures, and the speed at which permission changes are propagated all need to be verified during implementation. While the product architecture provides a direction, the actual results still depend on connection quality and process design.

V. Companies should first validate a closed-loop process rather than rolling it out company-wide all at once.

A more pragmatic pilot approach would be to select a process with clearly defined inputs, verifiable outcomes, and the ability to revert changes in case of failure. For example, start by having the agent collect data on abnormal orders and provide recommendations for handling them, then gradually expand to include low-risk operations.

Acceptance criteria can be divided into three categories: from a business perspective, processing time and rework rate; from a governance perspective, whether evidence is complete and approvals are valid; and from an operations perspective, whether the system can operate stably after rule modifications. Testing should also proactively cover expired data, revoked permissions, duplicate tasks, and tool failures—not just ideal inputs.

In its release notes, Sakana mentioned that experience implementing existing financial projects has driven the platform’s development, but these projects cannot be directly equated with the customer deployments Aquarium has already achieved. For the new platform, what is truly worth keeping an eye on is how subsequent specific use cases will demonstrate reusability, governance, and operational reliability.

Competition among enterprise agents will ultimately come down to “whether they can consistently handle a set of tasks for which they can be held accountable.” Aquarium offers this insight: Before pursuing greater automation, organizations must first establish a consistent understanding of data, rules, and accountability.

References

Release: Sakana Aquarium https://sakana.ai/aquarium-release/

Aquarium Product Overview https://sakana.ai/aquarium/