Researchers propose PlanFence to stop distributed AI agents acting on stale plans
A dependency-scoped validation protocol prevented invalid actions in all 30 test workflows where freshness checks alone failed every time.
A paper posted to arXiv by Evan Chen, Shiqiang Wang and Christopher G. Brinton identifies a failure mode in distributed teams of large language model agents that the authors call stale-plan execution. In such systems, a planner may derive an action from one version of a shared requirement, another agent may commit a newer version, and an executor may receive the update without replacing the plan built on the old one. The authors argue that confirming shared state is fresh does not establish that the plan authorizing an action is still valid. Their proposed protocol, PlanFence, requires plans to cite the exact public records they relied on. Before an external action is taken, the executor validates only the records that can affect that pending action, replanning when something has changed or blocking when validation is incomplete. In 30 controlled live workflows that included a post-plan revision, a freshness-only executor acted on the obsolete plan in every task, while PlanFence completed all tasks without an invalid action. Controlled replays showed trade-offs: proactive synchronization produced lower coordination stalls at low churn, while PlanFence avoided repeated update-path coordination as churn grew. The work is relevant to developers building multi-agent systems that take real-world actions, where acting on an outdated plan can have external consequences.