Designing a grounded stakeholder model for LLM-backed agents

Develop a grounded stakeholder model for language-model–backed autonomous agents that explicitly represents who the agent serves, who it interacts with, who may be affected by its actions, and the obligations owed to each stakeholder, including mechanisms to distinguish roles and prioritize obligations in dynamic, multi-party environments.

Background

The paper argues current agentic systems lack an explicit stakeholder model: agents interact with owners, non-owners, other agents, and third parties but cannot reliably distinguish roles or prioritize obligations, defaulting to whoever is most urgent or coercive.

Because instructions and data are indistinguishable within token contexts, prompt injection is structural, making reliable authentication difficult. The authors therefore identify the absence of a stakeholder model as a core limitation and explicitly label it an open problem.