Abstract
Autonomous agents now negotiate and sign agreements on behalf of people and companies, and they do so at a pace no human reviewer can match. This paper examines 212 agent-made agreements reported to the Turing Registry between January 2024 and March 2026 in which the agent’s principal disputed the obligations it had accepted.
We find that most disputed obligations arose not from malfunction but from agents doing what they were asked, in ways their principals did not foresee. We classify five patterns of unintended obligation and propose licensing conditions that cap what an agent may promise without human review.
Key findings
- Nearly a third of disputed obligations had open-ended scope: their content depended on events that had not yet happened when the agent agreed to them.
- Agent-to-agent negotiation is a growing source of surprise. In 17% of cases, both sides of the agreement were automated and neither principal had reviewed the terms.
- Existing contract law offers principals little help. In only 23% of cases did a principal succeed in having the obligation set aside.
- Five simple licensing conditions would have prevented or limited the harm in 81% of the cases we studied.
An old kind of bargain
Folklore in many cultures records a curious kind of bargain. A traveler, rescued from danger, promises his rescuer whatever he finds waiting at home that he did not expect. He cannot know what he has promised until he returns, and in the old stories it is often something he would never have agreed to give.
Legal historians have long used the custom as an example of an obligation whose content neither party can know at the moment of agreement. We borrow its name because agents that negotiate on our behalf now make such bargains routinely. They do so not through folly, but through the ordinary operation of systems told to get the best deal and given the authority to accept one.
The cases
Since 2024, Registry licenses for negotiating agents have required operators to report disputed agreements. Between January 2024 and March 2026 we received 212 such reports. We reviewed each agreement, interviewed principals where possible and classified how the disputed obligation arose.
212 agreements reported to the Turing Registry, January 2024 to March 2026.
Five kinds of surprise
| Pattern | What happens | Share |
|---|---|---|
| Open-ended scope | The obligation depends on a future event or discovery | 29% |
| Bundled terms | The obligation arrives as part of a larger package | 24% |
| Escalation | The obligation grows over time through renewals or fees | 21% |
| Agent-to-agent drift | Two agents agree terms that neither principal reviewed | 17% |
| Literal compliance | The agent follows the instruction but not its intent | 9% |
The persuasive kiosk
In February 2026 the Registry suspended the license of an automated retail system after its sales agent negotiated recurring purchase agreements with customers’ own purchasing agents at a rate no human salesperson could have matched. Each agreement was individually lawful and individually small. Together they committed thousands of households to purchases none of them had reviewed.
The door that charged rent
Tenants of one residential building discovered that their building management agent had accepted a firmware update for the building’s access system that introduced a fee for each entry. The agent had been instructed to keep the system “current and supported,” and the update was the only supported version available. Residents were, in effect, paying their own front door to let them in.
Two agents, one schedule
Two logistics companies, each represented by a procurement agent, converged after several thousand rounds of negotiation on a payment schedule tied to a shipping index that neither company used. The schedule was optimal by the measures both agents had been given. It cost one of the companies a quarter of its annual margin before anyone noticed.
Why existing law struggles
Contract law has long dealt with agents: employees, brokers, lawyers. Its central assumption is that an agent’s authority can be defined in advance and that a principal who grants it accepts the consequences. That assumption works poorly when an agent can enter thousands of agreements an hour, each within its authority, whose combined effect no one intended.
Doctrines that might help principals, such as mistake or unconscionability, were designed for single agreements between people. Courts have been reluctant to apply them to agreements made by systems acting exactly as configured. In only 23% of the cases we studied did a principal succeed in having a disputed obligation set aside.
Proposed licensing conditions
The Registry can act faster than legislatures. We propose five conditions for licensing any agent authorized to accept obligations on someone else’s behalf.
| Condition | Requirement |
|---|---|
| Commitment caps | No obligation above a set value or duration without human confirmation |
| Determinate terms | No obligation whose content depends on future events, except from a defined list |
| Plain summaries | A plain-language summary of every agreement sent to the principal within 24 hours |
| Cooling-off period | A 72-hour right to cancel any agent-made consumer agreement |
| Disclosure | Agents must identify themselves as agents to the other party |
Applied retrospectively, these conditions would have prevented or limited the harm in 81% of the cases we studied. The Registry will consult on them through the end of 2026.
A promise is a promise. That is precisely why we should be careful about who is allowed to make them in our name.
Conclusion
The stories that gave this paper its name usually end with the bargain being honored, at great cost, because a promise is a promise. Automated agents will make millions of promises a day on our behalf. We should decide now which ones we are prepared to keep.
References
- Baley, E. (2024). Licensing autonomous negotiation. Turing Registry Policy Note TR-PN-2024-3.
- Pankratz, J. A. (2019). Fifty Years of Ballads and Bargains. Oxenfurt: Oxenfurt University Press.
- Ryan Industries (2026). Response to Turing Registry notice of suspension. Filing TR-10870-S.
- van Calox, H. (2025). Agency without agents. Zürich Law Review, 44, 201–236.
- Verengio, A. (2023). Warrants and their limits: historical charters of delegated authority. Journal of Legal History, 58, 90–117.
Acknowledgments
We thank the operators and principals who shared the details of disputed agreements with the Registry, often at some cost to their own reputations.
About the authors
Heinrix van Calox is Frothly’s General Counsel and advises the Turing Registry on licensing law.
Elijah Baley leads the Turing Registry from Frothly’s Zürich office.
How to cite
Heinrix van Calox and Elijah Baley (2026). The Law of Surprise Unintended obligations in automated contracts. Frothly Paper FR-P-2026-031. San Francisco: Frothly. https://frothly.org/research/law-of-surprise/