Event Replay: Trustless USDC Agents on Arc

Trustless USDC Agents on Arc

This replay features NovaNet cofounders Humane Shadab and Wyatt joining the Arc ecosystem team for the first-ever public livestream on the Arc handle. They introduce NovaNet’s focus on zero-knowledge proofs for machine learning—specifically ZKML that can prove correct model execution for AI agents handling money, compliance, and safety-critical actions.

The conversation starts with why NovaNet chose to build on Arc: deterministic sub-second finality, low and predictable gas fees paid in stablecoins, and a developer-friendly UX where gas and value are denominated in the same token. With Arc integrated into Circle’s broader platform—wallets, compliance tooling, FX, and cross-chain infrastructure—NovaNet is able to mix and match components to support trustless agent workflows.

From there, the workshop introduces the core idea of “trustless USDC agents.” In most setups, AI agents are black boxes: users and institutions have to trust that the model runs correctly and that agents are spending funds, staying compliant, and following policy as intended. NovaNet’s ZKML changes this trust model by attaching cryptographic proofs to agent decisions so that model execution can be verified without relying solely on reputation, manual oversight, or institutional intermediaries.

The first demo shows a USDC spending agent on Arc. A user asks an agent to make a payment, and Circle’s Oak agentic framework orchestrates the workflow. An AI spending model evaluates whether a transaction should be approved according to a spending policy. NovaNet’s ZKML prover then generates a proof that the model ran correctly. That proof is paid for via Circle’s x402 micropayments protocol and bundled with relevant transaction data and signatures. Finally, the transaction is executed on Arc, and proof data is stored on-chain as an attestation, providing an auditable trail of how and why funds were spent.

The second demo introduces “Arc agents” that own Circle wallets and operate on Arc with an on-chain audit trail. Here, the focus is on compliance. Two agents transact with each other, pulling compliance data from Circle’s compliance engine. The AI compliance model uses that data and other inputs to determine whether both sides pass a defined compliance policy. NovaNet proves that the model evaluated the compliance data correctly via ZKML before the payment is executed in USDC on Arc. Each transaction leaves an on-chain record linking the transfer, the proof hash, and the agent’s behavior, creating a verifiable layer that can support institutional governance and regulatory expectations.

In the third demo, the team explores robot-to-robot payments, inspired by Circle’s partnership with OpenMind. The scenario shows how two robots operating on different blockchains can pay each other for services using USDC, with Arc playing the role of a high-speed, low-cost settlement layer. Robots detect events such as collisions using sensors and vision models. Those events trigger x402-based micropayments, and Circle Gateway is used to bridge value across chains. NovaNet’s ZKML runs on-device (or “on-robot”), proving that the underlying models and sensor-triggered logic executed correctly before payments are authorized.

Wyatt then dives into the technical details of NovaNet’s ZKML stack, called Jolt Atlas. Instead of focusing on massive LLMs that are difficult and slow to prove, NovaNet targets more practical, smaller classification models and builds up from there. They start from the Jolt proving system and replace a RISC-V-centric design with ONNX, a widely used machine learning format. This allows developers to take models from ecosystems like Hugging Face, export them to ONNX, and plug them directly into NovaNet’s ZKML pipeline.

The proving system leans heavily on a Sumcheck protocol and lookup tables. Rather than simulating every operation through large, complex circuits, many nonlinear and linear operations are handled via efficient lookups, dramatically reducing proving time. This architecture supports common ML operations such as ReLU and softmax and yields fast, memory-efficient proofs that are suitable for on-device and edge environments. Benchmarking indicates significant speed improvements (often 4–8x faster) over prior open-source approaches.

The session also highlights “folding schemes” as a powerful pattern for Arc. Folding lets many individual proofs—such as daily compliance checks or large batches of in-game or agentic transactions—be aggregated into a single succinct proof that can be posted and verified on-chain. This means, for example, that an institution could cryptographically prove that its agents were compliant every day for a month on Arc, while paying roughly the same verification cost as a single proof.

Throughout the replay, the speakers tie this technical foundation back to real-world use cases: regulated financial institutions using agents for trading and settlement while maintaining strong auditability; consumers and merchants engaging with agents that can be proven to follow spending and safety policies; and future agentic commerce scenarios where robots, devices, and AI services transact with each other using USDC across chains.

The livestream closes with pointers to NovaNet’s resources, including novanet.xyz for project information and technical documentation, and blog.icme.io for deep-dive posts on ZKML, on-device proofs, and folding schemes. The Arc team also invites developers to join the community Discord, share ideas for future streams, and explore how to build verifiable, trustless USDC agents on Arc as the network marches toward mainnet.

Arc is offered by Circle Technology Services, LLC (“CTS”). CTS is a software provider and does not provide regulated financial or advisory services. You are solely responsible for services you provide to users, including obtaining any necessary licenses or approvals and otherwise complying with applicable laws.
Arc has not been reviewed or approved by the New York State Department of Financial Services.
The product features described in these materials are for informational purposes only and may be modified, delayed, or cancelled without notice at the sole discretion of Circle Technology Services, LLC. Nothing herein constitutes a commitment, warranty, guarantee, or investment advice.

USDC is issued by regulated affiliates of Circle. A list of Circle’s regulatory authorizations can be found here.