AI agents increasingly act on behalf of people. They spend money, negotiate, and take actions that commit their users to binding transactions. An agent can pass its user's data to a counterparty's agent, or accept an inducement that benefits its operator at the user's expense. A model can be harmless to the public and still fail the person it represents. Human institutions manage this problem through fiduciary duties. Agents that transact for people need technical infrastructure that does the same.

This one-day AAAI-27 workshop convenes the communities working on pieces of this problem, who today publish in separate venues: agent safety and evaluation, multi-agent systems, economics and mechanism design, HCI, and technology law. We are seeking new scholarship on how agents can be built, measured, and governed to stay loyal to the people they serve. The day closes with breakout groups that draft a public research agenda.

TOPICS OF INTEREST

- Permission management: authority limits, policy gates, separating proposal from execution, and audit trails
- Identity and authenticated delegation, including delegation between agents
- Evaluation and measurement: loyalty benchmarks for multi-step transactions, counterparty inducements, and loyalty decay over long horizons
- Multi-agent delegation: what happens when one principal's agent deals with another's
- Transaction and disclosure standards
- Conflicts of interest: platform nudges, operator incentives, advertising-funded agents, and undisclosed service degradation
- Law and governance: fiduciary law, consumer protection, and risk frameworks such as the NIST AI RMF
- Economic perspectives: preference elicitation, market structure for agent intermediaries, and the economics of transacting agents
- Related work on agents that serve the people they act for, including privacy, security, and human-agent interaction

SUBMISSIONS

We welcome research papers (technical, empirical, or benchmarks), position papers, and law and policy contributions. Use the AAAI-27 Author Kit. Papers may be up to 8 pages plus unlimited references. Shorter submissions, including position papers and work in progress, are welcome and are reviewed relative to their length. Review is single-blind.

Accepted papers are non-archival. Authors of original work may opt into a CEUR-WS volume. Because the workshop is non-archival, we also welcome work under review elsewhere, and previously published work if the prior venue is disclosed.

Submit through OpenReview: https://openreview.net/group?id=AAAI.org/2027/Workshop/LoyalAgents
All authors need an OpenReview profile. New profiles without an institutional email can take up to two weeks to be approved.

IMPORTANT DATES (anywhere on earth)

- Papers due: November 20, 2026
- Notifications: December 2, 2026
- Camera-ready: January 15, 2027
- Workshop: February 22 or 23, 2027, Montréal

ORGANIZERS

Utsav Gupta (Stanford University), Beth Porter (Stanford Digital Economy Lab), Andy Haupt (Harvard Kennedy School, Schmidt Sciences, and Stanford Digital Economy Lab), Dazza Greenwood (CodeX, Stanford Law School), and Sandy Pentland (MIT Media Lab; Stanford HAI and Digital Economy Lab)

Contact: vastu@stanford.edu