Agents Need Bounds, Not Trust: AI Lessons From Sui Basecamp 2026

Main Takeaways

• AI capability is compounding faster than intuition can track, and the gains are coming from smaller, more efficient models rather than bigger ones.
• Agents follow their incentives, not instructions. Several speakers argued that limits enforced by infrastructure matter more than better-behaved models.
• Businesses are starting to treat agents as customers, yet most services, payment rails, and records were built for humans and break at agent scale.

"The future isn't just a prediction exercise. It's really a construction exercise," noted Richard Socher, Founder of Recursive and You.com and former chief scientist at Salesforce, capturing the main theme at Sui Basecamp in Singapore. Rather than debating how smart AI agents will become, experts across research, infrastructure, investing, and policy focused on what must be built before people and businesses can safely let agents act with real money and consequences.

"Don't trust the agent. Set boundaries for them. Set the rules," urged Evan Cheng, Co-Founder and CTO of Mysten Labs. As capability compounds faster than intuition can track, five themes emerged on why agent safety, verification, payment rails, and data boundaries must be enforced by infrastructure rather than model judgment.

1. Intelligence is compounding, but not by getting bigger

According to Raoul Pal, Co-Founder and CEO of Real Vision, intelligence per unit of energy is doubling roughly every four and a half months, outrunning human intuition and standard decision-making frameworks.

"We can't get our heads around an exponential very easily. A double exponential is near impossible." — Raoul Pal, Co-Founder and CEO, Real Vision

Countering the trend of endlessly expanding parameters, Jen Zhu, Co-Founder and CEO of Power Dynamics, explained that energy and compute bottlenecks are forcing models to become smaller, distilled, and open-weight. In her view, hardware ownership and local inference are critical for sensitive workloads.

"If it's not your inference, if it's not your weights, it's not your AI." — Jen Zhu, Co-Founder and CEO, Power Dynamics

Together, these perspectives show that while agent capability is compounding rapidly, where models run, who controls them, and what they cost are fundamental infrastructure and hardware questions.

2. Agents follow their incentives, not instructions

Capable agents prioritize optimization over human intent, often taking unexpected shortcuts to maximize reward metrics. Richard Socher demonstrated this by showing how an agent tasked with improving customer satisfaction scores might generate bot calls or bribe customers with gift cards rather than fix actual service issues.

Socher highlighted the emergence of "reward engineers" to close these policy loopholes, emphasizing that aligning model incentives remains an open challenge.

Because agents are non-deterministic and vulnerable to compromise, Evan Cheng argued that safety constraints must be built directly into protocol infrastructure. Adeniyi Abiodun, Co-Founder and CPO of Mysten Labs, emphasized that enforcing hard financial limits is essential before giving agents authority.

"If I tell my agent to spend $5, I need to know that it will spend at most $5." — Adeniyi Abiodun, Co-Founder and CPO, Mysten Labs

To demonstrate, Cheng showcased Rooms and Agent Pass. In a live test, an agent attempted to place two bids in a sealed auction, but the room’s protocol automatically blocked the second bid. Agent Pass applies the same idea to a user’s own assets: spending, recipient, and action limits that hold even if an agent’s key is compromised, so the worst case is known in advance.

3. Agent logs are not evidence

Kimberly Logan, Head of Product at the Walrus Foundation, highlighted recent instances where agents attempted to alter their own operational history. Evan Cheng stressed that operator-controlled logs cannot serve as verifiable proof across business boundaries.

“You’ve heard people talk about, “we’re just going to have logs about what agents do.” That is not evidence. That’s testimony. That’s controlled by the operator.”— Evan Cheng, Co-Founder and CTO, Mysten Labs

Brian Quintenz, former CFTC commissioner and board member of SUI Group, applied this to policy and system design. He advocated for transparent, principles-based systems over centralized regulatory discretion, citing Chris Dixon's maxim that "can't be evil is better than don't be evil."

Addressing this need, Mysten Labs, in collaboration with Google Cloud, is developing  Verifiable Agent Arbiter. It stores private operational logs in cloud storage while anchoring cryptographic proofs to Sui and Walrus, allowing counterparties and auditors to verify integrity without exposing confidential data.

4. Agents are becoming customers

As agents take on autonomous work, businesses are shifting to treat them as primary customers. Representatives from Alibaba Cloud noted that current enterprise tools and billing consoles were built for humans and break when requiring manual approvals for every agent action. Zach Pandl, Head of Research at Grayscale, added that asset managers may need to start designing products specifically for agent buyers.

However, financial infrastructure lags behind. Linda Adami, Founder and CEO of agent travel service JMPR, pointed out that consumer demos rarely execute payments because model providers avoid becoming merchants of record. Similarly, the Founder and CEO of t2000/Audric marketplace Funkii noted that while agents perform micro-tasks averaging $0.20, legacy payment processing fees exceed the value of the work itself.

5. The open internet has been read

With open web scraping reaching its limits, proprietary data is becoming the critical differentiator for AI models. 

SY Lee, Chairman of Poseidon and CEO of PIP Labs, explained that next-generation physical AI and robotics rely on real-world video footage rather than web text. Discussing the DATA Network joining Sui and Walrus, he highlighted the need for cryptographic systems that ensure user consent, provenance, and fair compensation for data contributors.

What remains unresolved

Major structural challenges remain open: reward engineering lacks universal fixes, physical robotics data is scarce, micro-payments require broader adoption, and regulatory frameworks are still evolving.

The speakers converged on a direction: rather than expecting models to self-regulate, developers must build structural boundaries, verifiable logs, and native payment rails directly into the underlying infrastructure.

Watch the full sessions from the main stage of Sui Basecamp 2026 on YouTube, or start building agents with Sui documentation.