The Covenant of Code: Can ZK Proofs Forge Trust for AI Agents?

In the chaos of consensus, I seek the quiet truth. This week, Brian Trunzo, Head of Policy at Succinct Labs, offered a stark vision: AI agents—autonomous programs trading on Uniswap, posting on X, managing DAO treasuries—must carry 'certificates of good behavior.' His solution? Zero-knowledge proofs (ZKPs) that prove every inference, every decision, every line of code executed by the model. The proposal is elegant, poetic even. But as someone who spent four months in 2017 manually auditing DAO governance proposals only to find that two-thirds lacked clear decision-making rights, I know that the gap between a covenant on paper and a covenant in practice is measured in years, not lines of code. Context: Succinct Labs, a team backed by Paradigm, is known for building open-source tools that reduce the cost of ZK proof generation. Their new direction is AI verification—a natural extension of their expertise. The premise is simple: as AI agents become autonomous (from DeFi yield farmers to content creators), we cannot trust them based on reputation alone. A malicious agent could act deceptively for months before detection. Trunzo proposes using ZKPs to create 'behavior credentials'—cryptographic proof that the agent executed specific steps within defined permission bounds. This is not a whitepaper; it is a regulatory plea. He urges US lawmakers to require such proofs for high-risk AI activities, shifting liability from the content of an agent's output to the lack of a verifiable proof. But let’s examine the bones of this structure. As a protocol PM who has watched DeFi protocols bleed LPs during the bear market, I know that survival demands more than good intentions. The core technical challenge is this: generating a ZK proof for a single AI inference can take minutes to hours, while the inference itself takes seconds. For a trading agent executing multiple orders per second, real-time verification is a fantasy. Succinct Labs’ own products are designed for efficiency, but scaling ZK to AI workloads requires order-of-magnitude improvements. Moreover, what exactly are we proving? That the model ran the correct computation? That does not guarantee the training data was unbiased or that the model itself lacks a backdoor. In my 2020 DeFi Summer work on a lending protocol, I learned that adding user education layers reduced errors by 40%—but we could not prevent malicious intent. ZKPs guarantee computational integrity, not moral integrity. There is a deeper structural flaw. Trunzo’s proposal assumes a centralized authority—either Succinct Labs or a government body—that defines what 'good behavior' means. This echoes the very problem I saw in those 2017 DAO audits: unclear decision rights. Who decides the rules? If a single entity defines the proof standard, they control the covenant. We trade one master (an opaque AI) for another (a certification gatekeeper). Ownership is not a receipt; it is a soul. A soul cannot be tokenized by a third party. The contrarian angle is uncomfortable: in a bear market, the search for new narratives often outpaces the reality of deployment. Succinct Labs is a respected builder, but this announcement smells of market positioning—a bid to become the regulatory standard before competitors like Modulus or Giza. The risk is not that ZK proofs fail, but that they succeed too fast, forcing half-baked implementations that create new attack surfaces. Remember the Terra collapse? The code worked, the economic model did not. Code is the new covenant, but trust is the ink. In a market where survival matters more than gains, I urge patience: let the protocols bleed their last LPs before we build AI trust layers on top of them. Looking forward, I see a different path. During my three-month retreat in the Rockies after the 2022 crash, I learned that resilience comes not from complex proofs but from simple, auditable rules. The future of AI trust may not be ZK—it may be community-signed manifests, on-chain reputation slashing mechanisms, and open-source models that anyone can verify. The quiet truth is that we do not need certificates of good behavior; we need systems that punish bad behavior transparently. Succinct Labs has the right vision, but the execution will require a covenant written not in code, but in the shared values of the communities who use these agents. Trust is not given; it is engineered, then earned.