Sui Basecamp 2026: What's Next?

Summary

At Sui Basecamp 2026, Evan Cheng outlined how Sui could provide security and coordination infrastructure for AI agents. His central argument was that as agents act faster and interact with other agents and businesses, human trust practices—such as informal permissions and blacklists—will not scale. Instead, agents should operate under explicit, enforceable limits, with independently verifiable records of their actions.

Key Concepts and Products

• On-chain policies and evidence: Cheng described using Move to define policies, permissions, and ownership rules, with Sui enforcing them. He distinguished verifiable evidence, which independent parties can check, from operator-controlled activity logs, which he called testimony. Policies and activity records can be anchored on-chain and independently replayed or checked.
• Sui Agent Pass: This permission system is intended to let an agent act with a user’s authorization while keeping limits and control with the user. Limits can cover actions and non-currency assets, not just coin allowances. The goal is to bound potential losses if an agent or key is compromised. Cheng said the current implementation is for vault accounts created for agents, with broader applications planned.
• Move across on- and off-chain workflows: The talk presented Move’s policy semantics as usable both on Sui and outside it, enabling businesses to automate internal workflows without extensive middleware. Cheng also previewed work on agent memory and a “memory compiler,” but did not explain how they would work.
• Agent Rooms: These were presented as controlled spaces where businesses and customer agents can interact under private rules and limits. Room activity and policies can be recorded on-chain, allowing agents to build reputations through verifiable interactions.

Live Demonstration

Cheng demonstrated a multi-agent store and sealed-bid auction. His Muse agent requested permission before bidding. The room rejected another agent after detecting that its declared bid differed from its submitted bid. Later, it blocked Cheng’s agent from claiming a second item because the rules allowed only one item per person.

The demo illustrated how predetermined protocols—not the agent or an operator—can enforce rules. The escrow deposit was described as occurring off-chain, while the policies and records were on-chain.

Outlook and Claims

Cheng argued that agent interactions will require composable policies, bounded permissions, and verifiable evidence. He predicted that more activity will move on-chain as agent use grows. He also said agent-based automation could simplify company workflows and reduce the risk of losing everything through a compromised key when limits are in place.

The talk mentioned a prospective verifiable-agent-arbitration effort involving Google, “Messina Lab,” and the Sui Foundation (names as captured in the subtitles). The initiatives discussed were planned for Q4 2026.

Reviews, Guides, or Tutorials

None. The talk included a product demonstration rather than a step-by-step guide.

Main Speaker and Source

Evan Cheng, identified in the video metadata as co-founder, CEO, and CTO of Mysten Labs. The presentation and live demo are the primary source.