AI News

Distributed LLM-agent teams can receive the latest shared facts while still executing a plan derived from an obsolete state. The paper calls this stale-plan execution: one agent may revise a requirement after a planner has formed a plan, while the executor receives the revision without replacing the old plan. It introduces PlanFence, a dependency-scoped action-validation protocol in which plans cite the exact public records used to create them. Before an external action, the executor validates only records that could affect that action, then replans when a relevant record changed or blocks when validation is incomplete. In 30 controlled live workflows containing a post-plan revision, a freshness-only executor used the obsolete plan in every task, while PlanFence completed every task without an invalid action. Controlled replay found that proactive synchronization caused less coordination stall when update churn was low. As churn increased, PlanFence avoided repeated coordination through the update path and avoided validating unrelated state as the shared keyspace grew. The authors frame these as controlled safety and systems-cost findings, not evidence of broader task-accuracy gains.