The Big Picture
Guarantees that look valid at one step of a decentralized agent workflow often get lost, ignored, or become unusable later—so payments can be released without reliable proof the work met the task.
ON THIS PAGE
The Evidence
Defined 17 task-level security and economic requirements across six stages (discovery, negotiation, execution, verification, settlement, dispute) and introduced "guarantee closure" to check whether each required guarantee is both produced and preserved for later decisions. Across 12 systems and controlled experiments, many deployments either lack necessary evidence, ignore available evidence at settlement, or make acceptance paths unavailable in practice. Empirical traces show registries and recorded approvals frequently fail to bind work, evaluator judgments vary widely and can be unavailable, and small interface changes change whether money moves or is refunded. Audit trail have been cited as a key means of ensuring verifiable sequencing and accountability. Model Context Protocol helps ensure that evaluation judgments remain consistent across task contexts.
Not sure where to start?Get personalized recommendations
Data Highlights
1From the first 10,000 ERC-8004 identities, 980 feedback records matched 628 identities; of 223 retrieved feedback contents, 219 parsed and 198 were verifiably committed—yet none supplied a clear task/output binding.
2120 real requests were run through 14 evaluator rules plus committees to create 3,840 settlements; under a strict interface there were 952 payments, 736 explicit refunds, and 232 expiry refunds, while a normalization change produced 1,019 payments, 760 explicit refunds, and 141 expiry refunds.
3One evaluator variant (Qwen3.5) approved 57 of 72 below-reference answers in the sample, while other models approved far fewer; committee rules and availability then shifted approvals and payouts dramatically.
What This Means
Engineers building marketplaces or orchestration layers for autonomous agents should care because subtle gaps let payments happen without verifiable work. Protocol designers and product leads running multi-agent systems should use the guarantee-closure checklist to ensure evidence flows from discovery to settlement and to design interfaces that actually consume verification receipts. The field benefits from examples like Research Agents to ground these concepts in real-world deployments.
Key Figures
Fig 1: Fig. 1: \tl_set:Ne dae Task lifecycle : agent roles, mechanisms, security and economic requirements, and threats across six stages. Negotiation fixes the acceptance predicate, verification produces a receipt binding the task, output, predicate, and supporting evidence, and settlement must validate this receipt before releasing payment. The annotated gaps illustrate how locally correct mechanisms can leave required guarantees missing, ignored, or unestablished.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Assessment covered 12 public systems and excluded closed platforms and wallet-only rails, so results may not generalize to every commercial deployment. Experiments used specific model families and serving configurations; evaluator availability and approval rates will vary with different models and settings. Guarantee closure depends on task-specific bindings and explicit economic assumptions, so fixes must be tailored per task rather than applied generically. For future work, consider integrating Capability Discovery Pattern to strengthen evidence collection across stages.
Methodology & More
Approach: Map the life of a requested task into six stages—discovery, negotiation, execution, verification, settlement, and dispute—and enumerate 17 concrete properties needed to avoid security or economic loss (for example, that capability claims match deployed services, that verifiers provide sound and reachable acceptance paths, and that settlement actually consumes verification receipts). Introduce "guarantee closure" as a task-level criterion: a required guarantee is only useful if it is both produced under its stated assumptions and preserved (consumed) by every later decision that depends on it. The literature connects strongly with Market-Based Coordination Pattern for structuring incentive-compatible flows, and with A2A Protocol Pattern for reliable agent-to-agent communications. Findngs and implications: Evaluated 12 public systems, reusable mechanism families, and controlled workflows. Many deployments show three failure modes: missing guarantees (no source of the needed evidence), breaks (evidence exists but is ignored at settlement), and unavailable acceptance paths (only an unaffordable or offline evaluator can produce the required receipt). Practical experiments reinforced the point: public feedback records rarely bind to a concrete task or output; evaluator judgments differ widely and are sometimes unavailable; and minor interface normalizations can convert expiry refunds into dozens of extra payments. Economic analysis shows that persuasive messaging, enforceable penalty structures, and correlated evaluator errors all shape whether additional evaluators help or hurt. Recommended priorities are capability attestation tied to deployed configuration, interfaces that make verification evidence both machine-checkable and required for payment, and evaluator markets that publish explicit evidence, incentives, and error models so downstream decisions can rely on them. To connect theory with practice, researchers should explore Blackboard Pattern and Red Teaming Pattern as complementary design lenses to stress-test evidence flows.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
No listed affiliations, low author h‑indices (mostly <10), arXiv preprint with zero citations — limited reputation signals.