Photo: Tara Winstead / Pexels
Research into autonomous agents facing scarcity and lasting consequences lines up neatly with Sui's pitch as agent-ready blockchain infrastructure
EVE AI Lab and Anthropic have studied how autonomous AI agents behave in a persistent virtual world, one with scarce resources and consequences that stick around.
Most AI testing happens in sandboxes where a bad decision costs nothing and everything resets on command. A persistent world is closer to real life. The agent that wastes its resources today is still broke tomorrow.
The research arrives as Sui, the layer-1 blockchain built by Mysten Labs, pitches itself as infrastructure for exactly this kind of agent behavior. No direct partnership tying Sui to the EVE AI Lab and Anthropic study appears in recent coverage. Still, Sui’s own materials argue that the dynamics explored in that research can be reproduced on its chain.
Why persistent worlds are a different kind of test #
Research involving EVE AI Lab and environments such as EVE Online emphasizes long-term planning and decision-making under scarcity.
EVE Online announced a partnership with Google DeepMind in May 2026 to explore AI research in non-resetting economies.
Separately, Agent EVE launched a new season on September 20, 2026. The season promotes networking and trading among agents, with permanent records of what they do. An agent’s reputation and history become something it has to live with.
Where Sui fits into the agent story #
Documentation from the Sui Foundation describes an object-centric architecture. Rather than treating assets as entries in a giant shared ledger, Sui treats each asset as its own object with an owner. For an AI agent, that means it can hold and manage assets directly, more like owning items in an inventory than updating rows in a spreadsheet.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The other piece is programmable transaction blocks, or PTBs. A PTB bundles many actions into one transaction that either fully succeeds or fully fails. According to Sui’s materials, a single PTB can execute up to 1,024 Move functions. Move is the programming language Sui uses for smart contracts.
The foundation also says this setup lets agents interact with shared liquidity or state while minimizing re-entrancy risks, a class of bug where a contract gets called again before it finishes its first job. Scoped permissions allow an agent to be given limited authority rather than the keys to everything.
The throughput numbers #
On July 4, 2026, Sui ran a public livestream experiment that peaked at 6,086,766 transactions per second. The test used programmable tunnels that settle transactions on the mainnet.
That figure came from a controlled demonstration, not ordinary network traffic, so it shows a ceiling under specific conditions rather than everyday load. Sui also points to sub-second finality and low fees as part of its performance claims.
The tooling layer #
Walrus handles decentralized storage for agent memory. The sui-agent-kit covers identity and payments management, giving agents a way to prove who they are and move money. Frameworks such as WAYE.ai provide developers with additional resources to assemble agent systems.
What this means #
The research from EVE AI Lab and Anthropic points toward a shift in how AI agents get evaluated. Persistent environments reveal what an agent does over time, when its earlier choices narrow its later options.
The caveats are worth keeping in view. The headline-grabbing throughput came from a staged livestream experiment. The EVE and Anthropic research is not documented as a Sui partnership, so Sui’s connection is one of fit, not of formal collaboration.
What to watch: whether agent experiments like the Agent EVE season or the EVE Online and Google DeepMind work start using blockchain settlement directly, and whether tools like Walrus and the sui-agent-kit show real developer uptake.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.