A Reference Architecture for AI Agents on Blockchain Infrastructure: Identity, Policy, Payments, and Custody as Composable Primitives

Abstract

Autonomous AI agents are rapidly being deployed as economic actors capable of discovering services, negotiating terms, executing transactions, and custodying value on blockchain infrastructure. Existing wallet, identity, and payment models were designed for human principals operating with full discretionary authority and offer few primitives for the bounded, delegated, and revocable authority that agentic operation requires. This paper proposes a layered reference architecture for AI agents on blockchain infrastructure, organized around four composable primitives — verifiable agent identity, bounded policy, constrained-authority custody, and intent-based payments — and unified by an intent based execution model in which identity is foundational. The architecture is grounded in emerging standards (ERC-8004, x402, ERC-4337, A2A, MCP, AP2) and mapped across heterogeneous settlement environments including the Ethereum Virtual Machine, Solana, the Canton Network, and XRPL. A case study of two parallel Canton initiatives — the proposed p402 trust-layer registries and the proposed Canton x402 facilitator — illustrates how the architecture's L1 identity / L4 payments separation manifests in practice and how privacy-preserving institutional ledgers reframe the agentic identity problem. A worked example using tokenized accounts receivable shows how the primitives compose under regulatory and operational constraints. The paper concludes with a discussion of audit, revocation, and open problems including agent-to-agent dispute resolution, cross-domain reputation portability, and the governance of solver networks.

Keywords