AI agents are taking on more autonomous roles, acting for humans or other agents in tasks such as buying services, sharing information, and delegating work. Their behavior is supposed to be governed by instructions, policies, or agreements, but these are typically expressed as plain text.
The new preprint argues that text alone cannot ensure an agent honors its commitments. To address this, the authors propose "proof-carrying contracts" — a mechanism that makes commitments binding in a way that agents cannot break. The abstract does not detail the technical construction, but the title suggests that contracts carry proofs that can be verified.
This could be a step toward more trustworthy autonomous agents, especially in commercial or delegated settings where enforcement matters. Since this is a single preprint, the ideas have not yet been peer-reviewed or compared with alternative approaches.