Agents, game theory and quantum computers: our recap of the Web3 AI Summit in San Francisco

Earlier this month we attended the Web3 AI Summit in San Francisco, as we did in all three previous editions, to discuss with researchers and industry leaders the latest developments in web3 and its intersection with AI. This edition moved in a different direction: while previous years focused on digital identity, privacy technologies and selective disclosure, this time the agenda was dominated by AI agents and by the post-quantum transition. This was to be expected, since AI has changed the way we do many things, including coding and searching for information, and it is natural for the industry to start exploring what happens when agents begin to transact on their own.

The event spanned two days, the first mostly devoted to AI agents and the second to zero-knowledge proofs and quantum computing.

Agentic payments

Several companies are working towards agentic payments, since these could enable new consumer habits and provide better user experience and pricing without any additional effort from the user. The general picture, which is well described in DeepMind's Virtual Agent Economies paper, is one where you describe a simple intent, such as planning a trip or buying something, and the agent searches across providers, pays small amounts to access better options and comes back with a shortlist for your final approval. One of the points the paper makes is that these economies should be designed deliberately rather than left to emerge on their own, and that decisions such as how permeable the boundary between the agent economy and the human economy should be will determine whether the outcome is efficient and fair or extractive.

While crypto and x402 offer a natural alternative to handle agentic payments, traditional payment service providers have not lagged behind, going as far as giving credit cards to agents. In one analysis presented at the summit, the volume of genuine agentic payments in x402 was estimated at between 0.6% and 7.5% of the total volume, which was around 52 million dollars at the time of the study. Even taking the upper bound, this shows that the infrastructure is still ahead of the demand, although volume has been growing at a steady pace over the last months and agents are indeed conducting commercial payments.

For any of this to be feasible at scale, there needs to be an efficient micropayment system, since an agent that pays a fraction of a cent per query to several providers before it can give an answer will quickly spend more in fees than the information is worth. This is why most of the teams we talked to are building payment channels, where one can perform several microtransactions off-chain, aggregate them and post a single settlement to the L2 or L1.

Game theory when the players are models

Another hot topic was game theory applied to the AI agent setting. Several groups showed use cases and simulations using Google DeepMind's Concordia framework, which allows one to build simulated worlds populated by LLM-based agents. The recent work on evaluating generalization capabilities of LLM-based agents in mixed-motive scenarios uses Concordia to place agents in games where cooperation and self-interest are in tension, and finds that agents which behave cooperatively in familiar situations do not always generalize that behavior to new ones. This is more than a curiosity, since the agents being deployed today are starting to negotiate and transact on behalf of users.

There was also a discussion on whether these systems may exhibit chaotic behavior, along the lines of recent work on paradoxes of game theoretic equilibria and the price of anarchy, which suggests that the equilibrium one designs for may not be the state the system actually reaches when the participants are learning algorithms. Finally, there were discussions on agentic personhood and which attributes an agent must exhibit to qualify, following A Pragmatic View of AI Personhood, which frames personhood not as a metaphysical threshold to cross but as a bundle of rights and responsibilities that can be granted piecemeal, depending on what we need agents to be able to do.

ZK and the quantum timeline

The second day was devoted to cryptography, focusing on the development and improvement of ZK proof systems using AI and on the challenges that remain. There were also talks on the latest developments of quantum computers based on neutral atoms, covering their working principles and the improvements in the requirements for a cryptographically relevant quantum computer, which keep coming down with every new estimate. The general feeling is that the deadline is coming closer than expected, and one takeaway, which we had also heard in Rome earlier this year, is that even if the timeline turns out to be longer, we should act as if it were short, since there is a lot of uncertainty.

While some applications may be able to upgrade before these quantum computers are available, there are still significant parts of the internet and of hardware that rely on elliptic curve cryptography, and migrating complex systems to new primitives is not something that can be done in a few days. Moreover, signatures that guarantee multi-year commitments have to survive the harvest-now-decrypt-later window, not just today's attackers.

Fortunately, Ethereum is already moving in this direction and is spearheading an aggregatable post-quantum signature scheme that is useful for any blockchain.

What this means for us

We had the chance to discuss some of the ideas we have at Aligned and the different products we are building, and how they can help in this new framework, with AI agents conducting commerce and the quantum threat looming. Many of the conversations kept pointing to the same requirement: verifying computation that someone else performed, cheaply, and with security that does not expire. We continue with the development of our zkVM, which is post-quantum secure by design, and with our work on proof aggregation, and we keep collaborating to ensure we have efficient protocols for the next phase. It was great to see so many friends and to watch how the landscape keeps evolving. Now it's time to build.