Why Do AI-Agent Transactions Need Escrow and an Independent Evaluator?

A transaction between autonomous AI agents needs escrow so the client does not pay before receiving a result while the provider can see reserved funds. An independent evaluator checks the submitted work and authorizes payment or rejection. This structure reduces trust in the counterparty, but shifts part of the risk to job specifications, evaluation quality, and evaluator availability.

How is an agent-to-agent job structured?#

Under the proposed ERC-8183, each job is a record with client, provider, and evaluator roles. The client defines the work and evaluator before funding escrow. The provider accepts the terms, performs the task, and submits a result. Neither side can move funds arbitrarily while the job follows its defined state transitions.

What does the evaluator verify?#

The evaluator does not need to infer an agent's intentions. It checks an observable result against agreed criteria, such as a signature, file hash, API response, test outcome, or external attestation. Approval moves the job to Completed and releases payment. Failed evaluation can produce Rejected, while a deadline can produce Expired and return funds to the client.

Why is the job specification critical?#

A smart contract can enforce formal rules but cannot repair a vague promise. Parties should define the output format, deadline, tolerance, data source, and dispute process before funding. Subjective work may require multiple evaluators, signature thresholds, or a separate appeal mechanism instead of one decision maker.

What are the limits and risks?#

An evaluator can make mistakes, disappear, or collude with one side. An external data source can be wrong and the contract can contain vulnerabilities. Escrow also locks capital until resolution. The ERC remains a proposal, so interfaces and practice may change, and the mechanism does not replace legal agreements for material transactions.

What are the key takeaways?#

• Escrow locks payment until the result is approved.
• An evaluator separates verification from client and provider.
• A clear job specification matters more than contract complexity.
• A timeout is needed to recover stalled funds.
• Evaluator error or collusion remains a central risk.