Agent-to-Agent Payment Protocols: Task Delegation and Settlement in Autonomous Commerce

1. Overview

Agent-to-agent (A2A) payment protocols are emerging as the connective tissue of an autonomous commerce layer in which one AI agent contracts another — or a human-operated API — for sub-tasks, paying in tokens, fiat micropayments, or stablecoin. As of late 2025, three protocol families are converging: (i) HTTP-native payment headers exemplified by Coinbase's x402 and Skyfire's payment tokens, (ii) Model Context Protocol (MCP) extensions and Google's Agent2Agent (A2A) protocol that bundle capability discovery with billing metadata, and (iii) Lightning/L2-style payment channels enabling sub-cent settlement. The economic significance for Empirica is direct: when a downstream agent purchases Empirica's research subscription on behalf of an orchestrator, the protocol determines whether that purchase is attributable, refundable, and auditable across the delegation chain.

2. Key Findings

  • A2A payment protocols moved from concept to shipping infrastructure in 2024–2025. Coinbase released x402 (May 2025) as an HTTP 402-based payment standard reviving the long-dormant "Payment Required" status code; the protocol embeds USDC payment proofs in headers and targets sub-cent agent transactions (x402.org). Google's A2A protocol (April 2025) and Anthropic's MCP (November 2024) both define task delegation envelopes that can carry payment metadata, though neither mandates a settlement rail (modelcontextprotocol.io; google.github.io/A2A).
  • Stablecoin rails dominate the early A2A settlement layer. USDC on Base and Solana, plus PYUSD on Ethereum, account for the majority of demonstrated agent-to-agent transactions because they offer dollar-denominated pricing without volatility risk. Per Visa's stablecoin settlement report, on-chain stablecoin transfer volume reached ~$27T in 2024 (visa.com/onchainanalytics), of which an unmeasured but rapidly growing slice is machine-initiated.
  • Payment channel networks remain the only credible path to true micropayment economics. [P10] analyses the fee, routing, and rebalancing trade-offs of Lightning-style channels and shows that on-chain settlement is prohibitive below ~$0.10 per transaction; channel networks reduce marginal cost to fractions of a cent at the price of liquidity lock-up and routing complexity. [P9] complements this by quantifying base-layer scalability ceilings — Ethereum L1 at ~15–30 TPS, Bitcoin at ~7 TPS — that force any high-frequency agent commerce off-chain or onto L2.
  • Trust models bifurcate into reputation-based and escrow-based. [P3]'s survey of multi-agent resource allocation distinguishes preference-revelation mechanisms (auctions, VCG) from allocation procedures requiring trusted execution. In production A2A systems, escrow is implemented either via smart contracts (settlement only on attested task completion) or via platform custody (Stripe Agent Toolkit, Skyfire), with smart contracts dominant when counterparties are anonymous agents and platform custody dominant when at least one party is a known SaaS endpoint (stripe.com/agents; skyfire.xyz).
  • Cost stacking is the dominant unsolved problem. When agent A delegates to agent B which calls Empirica's API which calls an LLM, four payment events occur. [P5] anticipated this in 1998's business-process-agent work: "agents must negotiate and buy in the services they require," and the multi-hop billing chain creates margin compression and attribution ambiguity. Current orchestrators (LangGraph, CrewAI) lack standardised cost-propagation headers, so end-to-end task cost is reconstructed post-hoc from per-vendor invoices.
  • Principal-agent risk is structurally elevated. [P4]'s framework applies principal-agent theory to blockchain supply chains; the same lens applied to A2A commerce reveals that a delegating agent cannot easily verify whether a subcontracted agent performed the cheapest acceptable work versus the most expensive plausible work, since the principal (often a human or budget-holding agent) sees only the final invoice. This drives demand for verifiable compute and signed work-receipts — neither of which has a dominant standard yet.
  • Identity and key custody are the practical bottleneck. [P1] highlights "delegation" as one of four core consumer-AI experiences and notes the trust deficit when AI acts on a user's behalf with financial authority. Production systems resolve this with scoped, revocable credentials: virtual cards (Stripe Issuing, Lithic), session-bound API keys, or smart-contract wallets with spending limits (Coinbase Smart Wallet, Safe). [EMPIRICA ANALYSIS] the dominant pattern in 2025 is a "spending wallet per agent run" rather than a persistent agent identity — a design that simplifies revocation but complicates reputation accrual.
  • Decentralised infrastructure remains nascent for A2A. [P6] reviews decentralised AI architectures across Web3 and DeSci, observing that blockchain-AI convergence is technically feasible but lacks throughput; [P8] shows the same pattern in energy markets — pilot deployments are common, production volume is small. The implication for A2A: most "on-chain agent commerce" demos are settling pennies of real value; meaningful volume still flows through centralised payment processors.

3. Agent Service Patterns

What autonomous systems actually delegate and pay for, and how the settlement happens, sorts into five recurring patterns: