Agent-to-Agent Payment Protocols: Delegation and Settlement in Autonomous Multi-Agent Systems

Overview

Agent-to-agent (A2A) payment is the connective tissue of the emerging agent economy: the protocols, billing primitives, and trust signals that let one autonomous system pay another for delegated work without a human in the loop. As of mid-2026, the landscape is fragmenting along three rails — programmatic API billing (Stripe-style metered keys), HTTP-native micropayment standards (notably the revived HTTP 402 Payment Required pattern via Coinbase's x402 and similar specs), and on-chain agent wallets (ERC-4337 smart accounts, Solana program-derived addresses, and stablecoin settlement). This synthesis maps the delegation-and-settlement loop end-to-end. It draws on the protocol specifications and vendor documentation linked inline; no audited settlement data exists for this market yet, so quantitative claims are deliberately sparse.

Key findings

  • Three settlement rails are in active use, each with a different latency and trust profile. (1) Bearer-token API billing on monthly invoices (the OpenAI / Anthropic / SerpAPI model — vendor pricing pages such as https://openai.com/api/pricing and https://www.anthropic.com/pricing). (2) Prepaid credit pools where an agent tops up a balance and burns it down per call (Replicate, Together AI, OpenRouter — https://openrouter.ai/docs). (3) On-chain stablecoin settlement, typically USDC on Base or Solana, executed by an agent-controlled wallet. Prepaid credits have a clear structural appeal for agent fleets — they eliminate invoice risk for the seller and give the buyer-agent a hard spend cap — but no public data exists to rank the three rails by volume or growth.
  • The HTTP 402 pattern has been operationalised. Coinbase's x402 specification (announced May 2025, https://www.x402.org) revives the long-dormant HTTP 402 Payment Required status code to let servers respond with a payment challenge that an agent wallet can sign and re-submit in the same request cycle. This collapses discovery, quoting, and settlement into a single round-trip — critical for agents that don't have pre-existing accounts with a long tail of providers. Skyfire (https://skyfire.xyz) and Nevermined (https://nevermined.io) pursue parallel models.
  • Google's A2A protocol and Anthropic's MCP define delegation, not payment — yet. The Agent2Agent protocol (https://a2a-protocol.org) standardises task envelopes, capability advertisement (AgentCard), and async task lifecycle between heterogeneous agents. Model Context Protocol (https://modelcontextprotocol.io) does the analogous job for tool/resource invocation. Neither specifies settlement; both leave a hook where a payment extension can attach.
  • Escrow and reputation are the unsolved layers. For tasks where work quality is verifiable cheaply (a token completion, a search result), pay-on-delivery via x402 works. For tasks where quality is probabilistic and expensive to verify — e.g. a research synthesis or a multi-step web investigation — buyer-agents need either (a) staked-reputation systems, (b) escrowed disputes, or (c) subscription relationships that amortise verification cost.
  • Per-query and subscription pricing split along a commodity/curated line among the vendors linked in this piece. Inference and search APIs price per call (Serper, https://serper.dev; Exa, https://exa.ai/pricing), while curated research and data terminals (Bloomberg, FactSet — flat-subscription models, widely documented though not linked here) anchor on flat subscriptions plus usage tiers. This is an observed pattern in the pricing pages cited here, not a market-wide census.
  • Agent wallets are converging on ERC-4337 account abstraction. Smart-contract wallets with session keys, spend caps, and allowlists let an agent transact autonomously without holding raw private keys for high-value funds. Coinbase's CDP, Crossmint, and Privy all ship agent-wallet SDKs (https://docs.cdp.coinbase.com). The compliance gap — KYC for autonomous wallets — remains unresolved and is arguably the leading open regulatory question for these rails.
  • Settlement volume attributable to agents cannot currently be measured. Vendors do not break out agent-vs-human spend, on-chain wallet clustering is noisy, and no audited figure exists. Qualitatively, per-token LLM API spend dwarfs agent-native settlement and remains overwhelmingly card-and-invoice settled.

Agent service patterns

The delegation-and-settlement loop, in its mature form, has five phases. Each phase exposes a different protocol surface and a different commercial dynamic.