ERC-8183 Closes the Escrow Gap: Pay AI Agents Only After They Deliver
While x402, h402, and AP2 handle payment transport and authorization, ERC-8183 introduces an on-chain escrow lifecycle that locks funds, accepts work, and requires explicit evaluation—closing the conditional settlement gap that existing protocols leave open.
The AI agent stack has payment rails, identity registries, and communication protocols. What it lacks is a standard way to make sure an agent actually delivers before it gets paid. ERC-8183 closes that gap with a job-based escrow flow: funds get locked, work gets submitted, an evaluator explicitly completes or rejects the job, and an expiry condition refunds the client if no resolution happens.[^claim_1964]
Existing payment protocols only move money. Coinbase’s x402 repurposes HTTP 402 for instant stablecoin micropayments, but it’s purely a transport rail—it moves money without deciding whether that money should move only after work is verified.[^claim_1966] The h402 protocol authenticates transactions through HTTP 402, targeting crypto microtransactions at 0.25% fees versus Stripe’s 1.5–4.35%.[^claim_1967] Google’s AP2 adds cryptographically signed payment mandates and Verifiable Credentials—tamper-proof, portable objects that prove what an agent is authorized to spend, independent of the payment method.[^claim_1965] None of these protocols implement escrow or work verification.
On the identity and authorization side, ERC-8004 gives agents verifiable identities, reputations, and capability proofs on-chain, led by the Ethereum Foundation’s dAI team with Google, Coinbase, and MetaMask.[^claim_1969] A2SPA requires every request between agents to be cryptographically signed, logs all interactions for replayable audit trails, and enforces policy-aware agents with dynamically scoped and revocable permissions.[^claim_1975] These standards tell you who an agent is and what it can do. They don’t tell you whether it actually performed the work.
Verifiable compute layers like EigenCompute and ZKML TEEs push execution off-chain while maintaining integrity. EigenCompute uses cryptoeconomic incentives to guarantee expected execution for complex, long-running AI agent logic.[^claim_1968] ZKML TEEs combine Trusted Execution Environments for confidential, high-performance ML execution with zero-knowledge proofs that cryptographically verify ML-based decisions without revealing private inputs.[^claim_1973] These verify that computation happened correctly. They don’t handle the conditional payment lifecycle—lock, evaluate, release.
Communication standards like MCP and A2A solve interoperability but not trust. MCP connects agents to tools and data sources with over 10,000 public servers and 97 million monthly SDK downloads, but it doesn’t verify whether a task outcome was actually delivered.[^claim_1970] A2A standardizes cross-system agent communication with over 100 supporting companies, but it doesn’t hold funds in escrow or judge deliverable quality.[^claim_1971]
CIPS adds context-aware routing and WASM-based sandboxed interoperability across EVM, SVM, and Cosmos SDK, reducing integration complexity by 4.7x and achieving 83% latency reduction on BNB Chain.[^claim_1974] Vitalik’s framework proposes local LLMs and zero-knowledge proof payment mechanisms so users can access AI services without exposing identity or data.[^claim_1972] ERC-8211 targets multi-step, dynamic on-chain operations for AI agents in DeFi, fixing failing execution chains from naive multi-call flows.[^claim_1976]
ERC-8183 is the missing conditional settlement layer. It composes with identity (ERC-8004), payment rails (x402, h402, AP2), and verifiable compute (EigenCompute, ZKML TEEs) to create a complete lifecycle: an agent with a verifiable identity (ERC-8004) sends a signed request (A2SPA), payment is authorized via AP2 and transported via x402, the work is executed off-chain with verifiable integrity (EigenCompute), and ERC-8183 locks funds until an evaluator confirms delivery. The remaining trust gap is around credible evaluators—future combinations with ERC-8004’s reputation or restaked verification networks could close that loop.
Provenance ledger
13 claims web-citedEvery claim below cites a source URL, and each URL was checked for validity before publish. The excerpt shown is the researcher's own summary of the page — it is not re-derived from the source, so it is not a verified verbatim quote. Follow the link to confirm any claim against the original. Citation markers in the text jump here.
[1] ERC-8183 defines a job-based escrow flow where funds are locked, work is submitted by an agent, an evaluator explicitly completes or rejects the job, and an expiry condition can refund the client if no resolution occurs. web-cited
“ERC-8183 defines a job-based escrow flow: funds locked, work submitted, evaluator completes or rejects, expiry can refund client… It is the article’s hook because it targets the missing conditional payment / verification step.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[2] Google’s Agent Payments Protocol (AP2) uses cryptographically signed payment mandates and Verifiable Credentials (VCs) as tamper‑proof, portable, cryptographically signed digital objects to prove what an AI agent is authorized to spend, independent of the payment method. web-cited
“Agent Payments Protocol (AP2) is an open protocol developed by Google… The core innovation of AP2 is the Verifiable Credentials (VCs) system - tamper-proof, portable, cryptographically signed digital objects… Google’s Agent Payment Protocol frames payments around cryptographically signed mandates that prove what an agent was permitted to spend.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[3] Coinbase’s x402 protocol repurposes the HTTP 402 Payment Required status code to enable instant, programmable stablecoin micropayments over HTTP, acting purely as a payment transport rail that moves money but does not implement escrow or work verification logic. web-cited
“Coinbase's x402 protocol handles the payment transport layer, enabling automatic stablecoin payments over HTTP… x402… moves money, but does not decide whether money should move only after work is verified; It is the payment transport rail, not the escrow/adjudication layer.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[4] The h402 payment protocol, based on the HTTP 402 standard, authenticates transactions directly through HTTP 402 and targets crypto microtransactions with fees of 0.25% compared to Stripe’s 1.5–4.35%. web-cited
“The outcome is the h402 protocol, a revolutionary payment standard set to debut in 2025, based on HTTP 402—a previously unused web standard now revitalized for crypto microtransactions… The h402 protocol authenticates transactions through HTTP 402, drastically reducing fees to 0.25% in contrast to Stripe’s 1.5–4.35%.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[5] EigenCloud’s EigenCompute is a verifiable offchain compute service that allows complex, long‑running AI agent logic to execute off‑chain while maintaining on‑chain integrity, using cryptoeconomic incentives to ensure expected execution and verifiable results for AI workloads. web-cited
“EigenCloud today launched EigenAI and EigenCompute on mainnet alpha, two services that bring verifiable computing to AI… EigenCompute is a verifiable offchain compute service that allows developers to run complex, long-running agent logic outside of a smart contract while maintaining the integrity and security of the onchain environment… results from AI agents can now be verified, with cryptoeconomic incentives designed to ensure expected execution.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[6] The ERC-8004 Trustless Agents standard defines on-chain identity, reputation, and capability-proof registries for AI agents, giving agents verifiable identities and reputations, and is led by the Ethereum Foundation’s dAI team in collaboration with Google, Coinbase, and MetaMask. web-cited
“The first infrastructure node worth paying attention to is ERC-8004… As the identity, reputation, and verification standard designed for AI Agents by Ethereum, it is led by the Ethereum Foundation's dAI team, in collaboration with Google, Coinbase, and MetaMask… its core logic… aims to give AI verifiable identities, reputations, and capability proofs on-chain.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[7] Anthropic’s Model Context Protocol (MCP) has reached more than 10,000 public servers and 97 million monthly SDK downloads, providing a standardized tool/data layer for connecting AI applications and agents to external tools, APIs, and data sources but not verifying task outcomes. web-cited
“Anthropic's Model Context Protocol now runs on more than 10,000 public servers and pulls 97 million monthly SDK downloads, connecting AI applications to external tools and data… MCP… connects AI applications and agents to external tools, APIs, and data sources… Does not verify whether a task outcome was actually delivered; It handles the tool/data layer, not the trust layer around completed work.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[8] Google’s Agent-to-Agent (A2A) protocol, launched in April 2025 with 50 partners and later expanded to more than 100 supporting companies, standardizes cross‑system AI agent communication and coordination but does not provide escrow or quality judgment for deliverables. web-cited
“Google's Agent-to-Agent protocol launched in April 2025 with 50 partners and scaled to more than 100 supporting companies before moving under Linux Foundation governance… A2A… lets agents communicate and coordinate across systems or organizations… Does not hold funds in escrow or judge deliverable quality; It solves agent interoperability, but not conditional settlement.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[9] Vitalik Buterin’s 2026 AI strategic framework proposes using local large language models and zero‑knowledge proof‑based payment mechanisms as trustworthy AI interaction tools so that users can access AI services and make payments without exposing identity or original data. web-cited
“Vitalik Buterin published a systematic AI strategic framework in early 2026… First is the trustworthy AI interaction tools. He advocates using local large language models (local LLMs), zero-knowledge proof payment mechanisms, and other tools to allow users to access AI services without exposing their identities and original data.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[10] ZKML TEE architectures in Web3 combine Trusted Execution Environments for confidential, high‑performance ML execution with zero‑knowledge proofs that cryptographically verify ML‑based decisions without revealing private inputs, targeting private and verifiable compute for AI in 2026. web-cited
“That is the promise of ZKML TEE blockchain privacy in 2026: TEEs deliver confidential, high-performance execution with attestation, while ZKML delivers cryptographic verification of ML-based decisions without revealing private inputs.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[11] CIPS (Contextual Interaction Protocol Standard) defines a unified command schema (CIP-01 to CIP-09), context-aware routing for execution path selection based on gas fees and latency, and WASM-based sandboxed interoperability across EVM, SVM, and Cosmos SDK chains, with an 83% latency reduction in a BNB Chain case study and 4.7x reduction in integration complexity versus conventional methods. web-cited
“CIPS (Contextual Interaction Protocol Standard) – a paradigm shift… Unified Command Schema: Standardizing cross-platform instruction sets (CIP-01 to CIP-09) Context-Aware Routing: Adaptive selection of execution paths based on gas fees and latency Sandboxed Interoperability: WASM-based virtualization supporting EVM, SVM, and Cosmos SDK environments… This framework reduces integration complexity by 4.7x compared to conventional methods, particularly enhancing blockchain interoperability (BNB Cha
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[12] A2SPA (Agent-to-Agent Signed Payload Architecture) requires every request between agents to be cryptographically signed, logs all interactions for replayable audit trails, and enforces policy-aware agents whose actions need toggle-state permission plus identity match, with dynamically scoped and revocable permissions. web-cited
“That’s where A2SPA comes in: Signed Payloads — Every request from one agent to another is cryptographically signed Audit Trails — Every interaction is logged, replayable, and verifiable Policy-Aware Agents — Actions require toggle-state permission + identity match Modular Trust — Agent permissions are dynamically scoped and revocable in real time.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[13] The Decentralized AI (dAI) effort on Ethereum, including ERC-8211, focuses on improving UX for AI agents in DeFi by enabling multi-step, dynamic on-chain operations to be executed more reliably than naive multi-call flows, which often result in failing execution chains for complex agent actions. web-cited
“In April 2026, Biconomy, together with the Ethereum Foundation, pushed for ERC-8211 in the direction of improving UX, aiming to solve the most realistic bottlenecks for AI Agents in the DeFi world, such as on-chain complex operations often requiring multiple calls rather than a single call, resulting in multi-step, dynamic, and easily failing execution chains.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
Sources
- https://cryptoslate.com/ai-agents-can-talk-use-tools-and-pay-but-crypto-wants-to-control-the-escrow-moment/
- https://dev.to/czmilo/2025-complete-guide-to-ai-agent-payments-how-the-ap2-protocol-is-reshaping-intelligent-commerce-2imf
- https://finance.yahoo.com/news/crypto-protocol-could-power-1t-183705481.html
- https://x.com/eigenlayer/status/1973010370798518362
- https://www.binance.com/en-IN/square/post/313025497738578
- https://www.assuredefi.com/blog/zero-knowledge-tees-the-future-of-private-and-verifiable-compute-in-web3
- https://www.btcwbo.com/index.php/2025/05/14/decentralized-neural-framework-next-generation-ai-agent-ecosystems/
- https://www.linkedin.com/posts/rakeshgohel01_starting-2025-ai-agent-interoperability-activity-7330572530061103105-ssMh