Ethan C. Jackson, PhD

A recent paper from Google DeepMind authors “Virtual Agent Economies” sets out a vision for how the economic system that will enable and govern agentic interactions may spontaneously emerge, but how it can also be shaped by technological protocols and design choices. Their framework highlights sandbox economies, questions of permeability vs. impermeability (i.e. degree of separation/isolation from human economies), and the need for mechanisms like auctions, reputation systems, and hybrid oversight. I agree with much of this analysis and I am strongly in alignment with many of the recommendations, both technical and social, made by the authors. One of the conclusions of the paper is a proposal that changes (so, integration and adoption of agents and agentic primitives into permeable economies) should be done in limited, gradual rollouts, and only with the support and buy-in of all stakeholders. While I agree that this approach would have obvious benefits in terms of safety and economic stability, I simply do not think that it is realistic.

The web3 world is not waiting for safe or impermeable sandboxes to be built before deploying agents into real, connected economic environments. AI agents are already being deployed as autonomous actors in financial systems (i.e. autonomous DeFi agents) and projects like Virtuals Protocol have already raced ahead with the introduction of digital agent currencies and a related Agent Commerce Protocol. While these projects may not have a lot of mainstream visibility, the point is that web3 doesn’t tend to move slowly or cautiously. Rather, web3 tends to look a lot more like high-stakes creative destruction, where hard lessons are learned via rapid innovation in real economic environments. And I do tend to agree, after understanding a bit more about how web3 works, that some hard lessons, especially around adversarial robustness, can only be learned in the trenches. I think applies to agentic AI robustness (and alignment, steerability, etc.) as it does to any mutiparty environment.

Still, going back to the paper: communication protocols, credit assignment, identity, reputation, credentials verification, guardrails, and incentive mechanisms can all be (and are) built on blockchain rails and/or verifiable compute. I agree with the recommendations that these are all things that must be advanced, ideally through the efforts of many independent contributing organizations, to ensure that decentralized agentic AI alignment is not only possible, but actually likely to emerge. I applaud the acknowledgement that our current trajectory points to the emergence of a decentralized and bottom-up agentic economy and that inclusive, participatory alignment and steerability are more likely to be achieved if this view is adopted. Just as human economies are shaped by incentives, regulations, and creative destruction, agentic economies will evolve through trial by fire in open markets, and outclass systems that are centrally planned.

For better or worse, Web3 x AI is already applying that creative destruction. It’s high-risk, but also brutally educational. And if history is any guide, it’s this real-world experimentation, rife with motivated adversaries, that will teach us some of the hardest lessons in scalable agent coordination.

The lesson: agentic AI isn’t going to be rolled out slowly, safely, and under perfect oversight. It’s going to emerge in the open, messy, permeable world of Web3 and other digital environments. And that’s exactly why blockchain rails, verifiable compute, and decentralized economic infrastructure matter. They’re not optional, they’re the only viable technologies we have to ensure the emergence of aligned and steerable agentic economies.