When an AI assistant participates in a group conversation and makes a promise to one participant, the question of who can later cancel that promise is not obvious. A new paper on arXiv (2610.21203) introduces the concept of 'standing' to address this. The authors define standing as the legitimate authority to release a commitment made by an agent.
The proposed rule is straightforward: only the creditor—the person to whom the promise was made—or someone explicitly authorized by the creditor or by the operator's policy may release the commitment. This restricts cancellation to those with a direct stake or formal permission, preventing other participants from unilaterally undoing an agent's promise.
The paper frames this as a principle for multi-party chat systems, where an assistant serves a team. By clarifying who has standing, the work aims to make agent behavior more predictable and accountable in collaborative settings.