As autonomous AI agents take on more of the negotiating, buying, and selling in digital commerce, a fundamental question remains unresolved: who adjudicates when the parties disagree? Albert Castellana, Co-Founder and CEO of GenLayer Labs , has built financial and blockchain infrastructure throughout his career — including at StakeHound and Radix — and is now applying that experience to what he calls the “adjudication layer for the agentic economy.” GenLayer’s Intelligent Contracts allow agreements to incorporate natural language and real-world context, with a decentralized network of AI validators reaching consensus on outcomes through a model the company calls Optimistic Democracy. In this interview, Castellana explains why traditional smart contracts can’t handle ambiguity, how GenLayer’s validator and appeals process works, and where he sees the technology heading as AI agents become active participants in the economy.
Q1. Looking at your experience building projects like StakeHound, Radix, and others, what key challenges in traditional dispute resolution urged you to create GenLayer as infrastructure for the agentic economy
I’ve spent most of my career building financial and blockchain infrastructure , and one thing kept coming up: the technology could make transactions dramatically faster, but the moment something went wrong, we were still relying on systems that were built for humans.
That problem becomes much more obvious when you look at autonomous agents. In traditional commerce, if two parties disagree, you can call a lawyer, go to court, or negotiate. That can already be painfully slow and expensive. I f millions of software agents are transacting with each other at machine speed , you can’t send every disagreement to a human process, it would be unthinkable.
That was the insight behind GenLayer… We needed an infrastructure layer for the part of a transaction that happens after the happy path, when the parties disagree about what was promised, what happened, or what the evidence means. GenLayer is our attempt to build that adjudication layer for the agentic economy.
Q2. As AI agents have started transacting autonomously at scale, who resolves disputes when something goes wrong, and why is this becoming a critical bottleneck?
Today, ultimately, a human does, and in our view that can be part of the problem. An agent can negotiate a transaction, make a payment and execute an agreement without a person involved. But when something goes wrong, we’re generally pushed back into human arbitration, customer support, legal processes or centralized platform decisions.
That model doesn’t scale to an economy where software agents could be conducting millions of transactions continuously. The individual disputes may be small, but collectively they become a huge infrastructure problem. A large share of digital disagreements simply aren’t worth pursuing because the cost and time of resolving them exceed the value of the transaction.
The missing piece is an adjudication layer that operates at the same speed and economic scale as the agents themselves. That’s what we’re building with GenLayer.
Q3. Why are existing layers such as payments, identity, and execution still insufficient for machine-to-machine commerce without a dedicated adjudication layer?
Payments answer the question, “How do I transfer value?” Identity answers, “Who am I dealing with?” Execution answers, “How do I carry out the transaction?” None of those answers the question, “What happens if we disagree about whether the agreement was fulfilled?”
That’s a fundamental gap. An agent might pay for a service, receive something it considers defective, and have a completely different interpretation from the other party about what was agreed. You need somewhere for that disagreement to go.
That’s why we think adjudication belongs in the stack itself. The transaction shouldn’t only specify what happens when everything goes right. It should also specify what happens when the parties disagree.
Q4. How does the GenLayer function as the missing infrastructure that enables trustworthy agreements between AI agents?
GenLayer gives agreements a way to handle the part that conventional smart contracts struggle with: judgment. With Intelligent Contracts, you can define an agreement in a way that includes natural language, live web information and other non-deterministic inputs. When a decision needs to be made, a selected validator proposes an outcome and an independent committee evaluates it. They don’t need to produce identical words. The contract defines what counts as an equivalent outcome, and the network reaches consensus around that.
If there is a disagreement, the decision can be appealed and reviewed by a fresh, larger committee. So you’re not asking one AI to be the judge. You’re creating a decentralized process through which independent AI validators reach an agreed outcome, with an economic mechanism for challenging it. That’s the infrastructure we think agent-to-agent commerce needs.
Q5. In what ways do Intelligent Contracts enable GenLayer to handle real-world ambiguity and context that traditional smart contracts cannot?
Traditional smart contracts are extremely good at deterministic rules. If X happens, do Y. That’s a feature, not a flaw.
But real-world agreements aren’t always like that. They contain terms such as “substantially completed,” “reasonable quality,” “delivered on time,” or “meets the agreed specification.” You can try to reduce all of those concepts to rigid code, but eventually you’re trying to encode human judgment into a system that was designed to avoid it.
Intelligent Contracts approach the problem differently. They can interpret natural language, access live web information and perform non-deterministic operations, while the consensus layer provides a way for independent validators to assess those results.
That’s what makes them interesting. We aren’t trying to eliminate ambiguity from agreements. We’re building infrastructure that can actually deal with it.
Q6. How does the network’s consensus approach help independent AI validators reach reliable decisions on subjective outcomes at machine speed?
The key is that we’re not asking every AI to magically produce the same answer.
In GenLayer’s Optimistic Democracy, a leader proposes an execution result and a stake-weighted committee of independently selected validators evaluates it. For non-deterministic outputs, validators use the contract’s Equivalence Principle to determine whether the result is acceptable.
That distinction matters. If five models interpret a clause slightly differently, we don’t necessarily need five identical paragraphs. We need to know whether their conclusions satisfy the criteria defined by the contract.
And if someone believes the decision is wrong, they can challenge it. Appeals bring in fresh validators and larger committees, increasing the cost of sustaining an incorrect decision.
So the goal isn’t to pretend AI is infallible. It’s to create a decentralized process for reaching and challenging decisions without putting one model or one company in charge.
Q7. What advantages does this adjudication model offer for use cases in agentic commerce, prediction markets, or autonomous finance compared with conventional systems?
The common thread across these applications is that they involve decisions that are difficult to express entirely as deterministic code. Take agentic commerce. Two agents can agree on a service, payment and delivery terms, but someone still needs to determine whether the service actually met the agreement. In prediction markets, the question may not be whether a particular number exists in a database, but whether a real-world event satisfies the market’s resolution criteria. In autonomous finance, you can have agreements whose execution depends on interpreting external information or contractual conditions.
GenLayer brings those decisions into the same programmable environment as the transaction itself. That can make them faster, more automated and more scalable than routing every disagreement through a centralized platform or human process.
The important distinction is that we’re not saying AI is inherently better than human judgment. Instead we’re building a system for a class of disputes that needs to operate at machine speed and where the economics often don’t justify traditional adjudication.
Q8. Looking at current testnet activity and builder adoption, how is GenLayer progressing to become an essential infrastructure for the agentic economy?
The most important signal for us is that people are actually building and testing the infrastructure.
We’re already seeing thousands of decisions being processed across the network, more than 200 builders working on the protocol, a community of more than 80,000 people, and thousands of GitHub stars across the open-source codebase. Rally, one of the flagship applications built around the technology, has also generated more than 200 million impressions and reached more than 140,000 users.
But I don’t think the right way to measure this is simply by counting users or transactions. We’re trying to prove something technically difficult: that a decentralized network can coordinate independent AI systems around decisions that aren’t deterministic.
Our testnet progression is designed around proving those properties step by step. The goal is to arrive at mainnet with the protocol, validator economics and consensus mechanism having been tested in increasingly demanding conditions.
We’re still early. That’s important to say. The objective now is to turn that technical progress and builder activity into infrastructure people can rely on in production.
Q9. In the long run, what is your vision for GenLayer’s role in supporting scalable, enforceable machine-to-machine commerce as AI agents become more widespread?
I would say we’re moving toward a world where software won’t just execute instructions. It will negotiate, make decisions, buy things, sell things and enter into agreements on our behalf. If that happens at scale, we need to rethink what a contract means.
My long-term vision for GenLayer is that it becomes the neutral adjudication infrastructure underneath that economy. When two agents agree to something, they shouldn’t need to trust each other completely, and they shouldn’t need to involve a human every time they disagree. There should be a protocol they can both rely on to determine what the agreement means and whether it was fulfilled. That’s what I mean by the adjudication layer for the agentic economy.
We’re not trying to replace courts or claim that every dispute can be reduced to an AI decision. There will always be situations that belong in existing legal systems. The opportunity is to build the internet-native first layer for the enormous number of smaller, faster, machine-to-machine disputes that traditional systems were never designed to handle.
If agents are going to become economic actors, they need more than the ability to transact. They need a way to resolve disagreement. I think that’s a fundamental piece of infrastructure the internet is still missing.