When Handoffs Go Stale: Validity-Aware State Transfer for Multi-Agent LLMs

Multi-agent LLM systems rely on handoffs to continue tasks across specialized roles. To fit limited context budgets, existing methods summarize or retrieve task history, but may omit constraints, verification status, and dependencies even when relevant facts are retained. We formulate this gap as a receiver-conditioned state-projection problem and propose a training-free method for validity-aware state transfer. The method maintains a typed task-state graph that tracks provenance, evidence, and lifecycle changes, constructs budgeted capsules containing action-relevant state and its validity relations, and repairs affected state slots using execution feedback. Experiments on SWE-bench Lite show that our method achieves a 63.3% resolution rate, compared with 40.7% for vector memory, the strongest evaluated multi-agent baseline by task success, while reducing total tokens per resolved instance by 46.6%. Further controlled evaluations across software repair, knowledge analysis, and enterprise workflows demonstrate improved retention of critical state, identification of stale information, and recovery from handoff errors relative to summarization and retrieval baselines.