A design note released by the Sui Foundation on October 28 discusses how AI agents operating on the Sui blockchain can be constrained by enforceable permissions and verifiable records. The note emphasizes that agents need both sufficient authority to complete tasks and clear boundaries to prevent errors or misuse.
Permissions versus Instructions
The note distinguishes between simple instructions, such as telling software to stay under a spending limit, and system‑enforced permissions that prevent the agent from exceeding that limit regardless of its interpretation. Effective deployment, it argues, requires both a sensible objective and a mechanism that can reject unauthorized actions.
Atomic Transactions
Sui’s programmable transaction model allows multiple commands to be grouped into a single atomic transaction. If any command fails, the entire group is rolled back, avoiding partial asset changes. This can be useful for workflows where several on‑chain steps depend on each other, such as a purchase followed by asset transfer. However, atomic execution does not guarantee that the price, recipient, or external service is appropriate.
Delegation and Authority
When an agent delegates part of a task to another service, the design note stresses the need for visible, enforceable boundaries. Permissions may include a budget, approved purpose, expiry, and a limited set of counterparties, with the system checking these conditions at the point of authorization.
Verifiable Records
Shared on‑chain records can help participants reconcile actions and investigate disputes. While a successful transaction confirms that an on‑chain operation occurred, it does not prove off‑chain outcomes such as delivery of physical goods or invoice authenticity, which require separate evidence.
Practical Implications
The note concludes that developers should start with narrowly defined authority, enforce it at the system boundary, and make completed actions inspectable. Expansion of scope should only occur when additional controls and operational evidence support it. The overall quality of an AI agent depends on the complete application, including permissions, data sources, and handling of unexpected workflow deviations.
