An agent handling a business request needs to know which record is current and who can authorize the next step. Sakana Aquarium puts those questions into the platform’s design.
Sakana’s enterprise-agent announcement
Sakana AI announced Aquarium on October 23 as a platform for building and operating organizational agents. Its published design combines an ontology, governance controls, and managed agents.
An ontology defines data meaning and the relationships among records, rules, and concepts. Aquarium’s product page describes connecting existing data sources to that shared structure. The integration and synchronization method depends on the underlying system.
The announcement states Sakana’s architecture and product direction. Claims about safer operation or reduced implementation effort remain company claims until a deployment supplies evidence.
Define what an approval actually authorizes
The Aquarium product page describes workflows written in YAML, including human approval steps. Sakana also describes access controls and records of the evidence, decisions, and actions within a workflow.
A useful buyer question is how an approval relates to the state of the underlying data. If a person approves a request and the record changes afterward, the implementation must define whether the approval still applies. The public announcement doesn’t answer every operational detail.
That question gives a concrete starting point for a consultation. Ask the provider to demonstrate the specific approval rule using a disposable record and a documented expected result.
Start with one operation you can inspect
Consider a proposed stock-transfer request. An agent reads a stock record, proposes moving an amount between locations, and asks an authorized person to approve it. That example is a design exercise, not a claimed Aquarium customer deployment.
- Name the source of the current stock figure.
- Define which fields the agent may read and change.
- Record the amount and record version presented for approval.
- Specify what happens if the available stock changes before the action.
- Verify the final record and the action log after an approved trial.
A buyer can use the same exercise to ask about failure handling. Determine who can stop the workflow, correct an incorrect record, and trace an action back to the approval that allowed it. These are requirements to verify with the implementation team.
Access currently begins with a consultation
Sakana’s public page invites organizations to discuss their operations and implementation needs. It does not establish general self-service access, public prices, or a complete regional availability list.
The announcement also describes future work to incorporate feedback and update the ontology and agents over time. Those plans should remain separate from capabilities demonstrated in a current deployment.
For a prospective buyer, the useful next deliverable is a narrowly scoped demonstration with agreed records and outcomes. Request evidence for the operation you want to automate before extrapolating to an entire organization.
Related reading: Google’s announced agent-compute preview and separating telemetry from automated repairs.
Sources
Which operation in your business has a clear enough approval rule to use as a first demonstration?
Sources checked October 26, 2026. Prepared with AI-assisted research and drafting. Proposed checks are illustrative; MustHave.ai has not independently tested this system.
