PlanFence
A safety protocol addressing stale plan execution in distributed multi-agent systems. Through a dependency-scoped validation mechanism, it requires plans to explicitly annotate the shared records they depend on and verifies the validity of those records before execution, thereby preventing ineffective actions taken on the basis of outdated plans.
Timeline (last 90 days)
The arXiv paper "Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory" proposes the PlanFence protocol to address the stale-plan execution problem.
The core design philosophy of PlanFence is dependency-scoped validation, focusing only on the records that truly affect the pending external actions.
PlanFence requires that a plan must explicitly reference the exact public records on which it is based.
The PlanFence executor validates the records referenced by a plan before executing external actions, replans once if the records have been updated, and blocks if validation cannot be completed.
The researchers tested PlanFence in 30 controlled real-time workflows that included post-plan-generation revisions.
Executors that rely solely on state freshness acted on outdated plans in every task, with a 100% failure rate.
The PlanFence protocol completed all test tasks without producing any invalid actions.
When the frequency of shared state changes is low, proactive synchronization strategies result in lower coordination stalls than PlanFence.
The larger the system scale and the more frequent the changes, the more cost-advantageous PlanFence's dependency-scoped validation becomes.
The paper's authors emphasize that these results are safety and system cost results under controlled conditions, not general improvements in task accuracy.
1 more timeline events
All Facts (11)
The arXiv paper "Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory" proposes the PlanFence protocol to address the stale-plan execution problem.
50%UnverifiedThe core design philosophy of PlanFence is dependency-scoped validation, focusing only on the records that truly affect the pending external actions.
50%UnverifiedPlanFence requires that a plan must explicitly reference the exact public records on which it is based.
50%UnverifiedThe PlanFence executor validates the records referenced by a plan before executing external actions, replans once if the records have been updated, and blocks if validation cannot be completed.
50%UnverifiedThe researchers tested PlanFence in 30 controlled real-time workflows that included post-plan-generation revisions.
50%UnverifiedExecutors that rely solely on state freshness acted on outdated plans in every task, with a 100% failure rate.
50%UnverifiedThe PlanFence protocol completed all test tasks without producing any invalid actions.
50%UnverifiedWhen the frequency of shared state changes is low, proactive synchronization strategies result in lower coordination stalls than PlanFence.
50%UnverifiedThe larger the system scale and the more frequent the changes, the more cost-advantageous PlanFence's dependency-scoped validation becomes.
50%UnverifiedThe paper's authors emphasize that these results are safety and system cost results under controlled conditions, not general improvements in task accuracy.
50%UnverifiedPlanFence does not improve task completion correctness, but rather addresses coordination safety, ensuring that agents do not execute irreversible external actions when a plan has become invalid.
50%