RightSource: Joint Planning of Evidence and Authorization for Tool-Using Agents

Tool-using agents must determine not only whether an action satisfies its task requirements, but also whether the appropriate party has authorized it. These decisions become difficult when relevant information is incomplete, evidence comes from different sources, and confirmations apply only to particular actions. A successful factual check does not establish permission, while requesting every possible confirmation can introduce unnecessary interaction. We introduce **RightSource**, a framework for synthesizing verifiable decision policies that distinguish *factual evidence* from *action-scoped authorization*. RightSource represents uncertainty as a set of possible worlds and records confirmations separately according to their source and scope. It plans which observations and confirmations to acquire, allowing subsequent queries to depend on earlier outcomes. Before permitting an action, the framework checks that its safety conditions hold in every remaining possible world and that all required authorizations have been obtained. Within the specified finite model and registered sources, the resulting policies provide checkable guarantees for permitted actions. We evaluate RightSource on retail workflows adapted from τ-bench and prompt-injection scenarios from AgentDojo. The results show fewer retail queries with similar observed task-completion rates, and fewer argument-scope violations in the tested injection scenarios, supporting joint planning of evidence acquisition and authorization.