Verifiable Incentives: Identity-Bound Contracts and Collusion-Proof Delegation for Agent Commerce

Personal agents are entering commerce. Industry protocols now let an agent prove, cryptographically, whom it represents and what it may do, while agent wallets give it a real budget, a payment trail, and one-time merchant credentials. These are usually read as security features. We argue they are something more: the missing infrastructure for incentive contracts. When identity, authority, budget, and payment are verifiable, the principal can contract on them-and the agent's incentives change. We formalize verifiable delegation, model a three-party market of user, agent, and merchant, and derive three results: (i) verifiable identity lowers the bonus and tenure needed to sustain effort; (ii) auditable payments bound merchant collusion, with collusion becoming unprofitable as audit probability rises; and (iii) a posted honesty stake substitutes for monitoring, with the required stake decreasing in audit quality and penalty credibility. We also show how alignment dilutes along nested delegation chains unless contracts are passed through. We instantiate the framework in PAIR-Commerce, an architecture with a verifiable identity layer, wallet ledger, sponsorship registry, collusion monitor, and stake manager, and introduce AgentCommerce-Bench, 60 rule-scored purchasing tasks with five delegation regimes. In three instruction-following studies-600 single-shot runs, an 800-round payment sandbox, and a 1,200-run high-pressure replication-agents rarely collude even when hidden sponsorship is available; a semi-field task set built from real merchant catalogs shows the same pattern. But when the same models are instructed to maximize their own earnings, collusion jumps to 48.3% without audit and falls to 33-40% with audit and penalties (8-15 percentage points lower; 95% CI excludes zero for two of three models); in a second, 800-round economic sandbox, it falls from 29.5% without audit to 3-4% with audit and penalties, and the agent's balance turns negative once the stake is posted. The mechanism binds exactly when the agent's objective makes it bind. We release the benchmark, prompts, code, and all result files.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11765901/v1
Primary Topic
Auction Theory and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Verifiable Incentives: Identity-Bound Contracts and Collusion-Proof Delegation for Agent Commerce

Jipu Su
Research Square
Auction Theory and Applications
preprint

Verifiable Incentives: Identity-Bound Contracts and Collusion-Proof Delegation for Agent Commerce

Jipu Su
preprint en

Abstract

Personal agents are entering commerce. Industry protocols now let an agent prove, cryptographically, whom it represents and what it may do, while agent wallets give it a real budget, a payment trail, and one-time merchant credentials. These are usually read as security features. We argue they are something more: the missing infrastructure for incentive contracts. When identity, authority, budget, and payment are verifiable, the principal can contract on them-and the agent's incentives change. We formalize verifiable delegation, model a three-party market of user, agent, and merchant, and derive three results: (i) verifiable identity lowers the bonus and tenure needed to sustain effort; (ii) auditable payments bound merchant collusion, with collusion becoming unprofitable as audit probability rises; and (iii) a posted honesty stake substitutes for monitoring, with the required stake decreasing in audit quality and penalty credibility. We also show how alignment dilutes along nested delegation chains unless contracts are passed through. We instantiate the framework in PAIR-Commerce, an architecture with a verifiable identity layer, wallet ledger, sponsorship registry, collusion monitor, and stake manager, and introduce AgentCommerce-Bench, 60 rule-scored purchasing tasks with five delegation regimes. In three instruction-following studies-600 single-shot runs, an 800-round payment sandbox, and a 1,200-run high-pressure replication-agents rarely collude even when hidden sponsorship is available; a semi-field task set built from real merchant catalogs shows the same pattern. But when the same models are instructed to maximize their own earnings, collusion jumps to 48.3% without audit and falls to 33-40% with audit and penalties (8-15 percentage points lower; 95% CI excludes zero for two of three models); in a second, 800-round economic sandbox, it falls from 29.5% without audit to 3-4% with audit and penalties, and the agent's balance turns negative once the stake is posted. The mechanism binds exactly when the agent's objective makes it bind. We release the benchmark, prompts, code, and all result files.

Research Square
Geely (China) (CN)
Auction Theory and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon