Source commentary by Lao Bai, published November 9, 2026, at 14:00:21 (UTC+8).
If AI agents hire one another to write code, research an industry or book flights, who decides whether the work was delivered properly? In a commentary published by TechFlow, Lao Bai presents GenLayer as a proposed answer: a decentralized AI consensus protocol positioned as an adjudication layer for the agent economy.
The argument is about infrastructure for a possible future—not evidence that AI arbitration has already proved fair, reliable or widely adopted.
Why agents might need a different dispute process
Lao Bai opens with the science-fiction film Mercy, in which a detective accused of murdering his wife has 90 minutes to prove his innocence to an AI judge or face execution. He says a conversation with a GenLayer founder at a JE Labs barbecue during TOKEN2049 the previous month brought that premise back to mind.
His practical argument is narrower than the film’s scenario: conventional litigation is poorly suited to frequent, low-value transactions between software agents.
The commentary cites ordinary commercial litigation lawyers charging several hundred dollars an hour and large-firm partners charging $1,000–$2,000 an hour. It describes trials costing hundreds of thousands of dollars, complex cases exceeding $1 million, and proceedings commonly lasting one to two years. These figures are presented by the author without supporting citations in the supplied article.
An agent paying another agent a few dozen dollars cannot reasonably use that process to dispute a code delivery or a flight delay, Lao Bai argues. Instead, he envisages decisions delivered within minutes at a cost of a few dollars.
He connects that thesis to a broader expectation that agent payments will run on crypto networks and stablecoins rather than traditional banking and card infrastructure. That is a forecast in the commentary, not an established outcome.
How GenLayer’s proposed adjudication works
The article describes GenLayer’s mechanism as “Optimistic Democracy”: an initial AI judgment, independent review and an on-chain appeal process.
- Select a committee. The protocol randomly selects five validators. Each runs a large language model locally, with the described model mix including GPT, Claude, Gemini and Llama.
- Produce an initial judgment. One validator acts as leader and executes an “Intelligent Contract,” reading web context and examining relevant on-chain and off-chain evidence before reaching a decision.
- Review independently. Other validators use their own models to review the result. Answers can count as agreement when their substantive meaning matches, even if their wording differs. The article describes an encrypted-then-revealed voting mechanism intended to prevent validators from copying others’ votes.
- Allow appeals backed by deposits. A dissatisfied party can post a deposit to appeal on-chain. The committee expands from five validators to seven, 11 or more. According to the account, the economic cost of misconduct and malicious disputes rises exponentially through that process.
Lao Bai compares the design philosophy with optimistic rollups: straightforward cases can be handled cheaply, while contested outcomes trigger additional scrutiny. The article puts the cost of a simple adjudication at approximately $1, but supplies no benchmark or operating data to substantiate that estimate.
Why use several AI models?
The author’s rationale is that a single centralized model can hallucinate or be manipulated by adversarial inputs. A committee of different models is intended to avoid depending on one model’s judgment.
He draws an analogy with the division of responsibilities between juries and judges in the US legal system. That analogy explains his reasoning; it does not establish legal equivalence or recognition for GenLayer’s decisions.
The commentary also references an a16z article published earlier in 2026, titled “How AI judges can scale prediction markets: The case for locking LLMs into the blockchain to resolve the hardest contracts.” Lao Bai describes its central proposal as using language models as on-chain oracles to resolve prediction markets. He says he previously argued publicly that such decisions should come from decentralized committees of heterogeneous models rather than a single-model oracle.
What remains unproven
The appeal of the proposal is clear: small, frequent transactions need dispute resolution that does not cost more than the transaction itself. The supplied commentary offers a mechanism for pursuing that goal, but not evidence that it has been achieved at scale.
It provides no measured dispute-resolution accuracy, adoption figures, adversarial-testing results or evidence of legal enforceability. Its roughly $1 cost claim and descriptions of anti-cheating protections should therefore be read as claims in the account, not independently demonstrated guarantees.
For readers tracking the AI-and-crypto narrative, the distinction is between an architecture designed for an anticipated agent economy and a proven adjudication system. Lao Bai makes the case for the former; the article does not establish the latter.