What problem does it solve? Hierarchical multi-agent systems waste tokens by replaying the full orchestrator trajectory into every worker call, while summarization adds latency and loses information. This Skill explains how to transfer orchestrator state to workers at the representation level by compacting the worker model's KV cache with Attention Matching. ## Core Features & Use Cases - Task-Guided KV Compaction: Score trajectory positions using queries from the current worker prompt and retain only the positions the worker actually attends to. - Shared Global Mask with MAD Thresholding: Aggregate attention scores across layers and heads into one mask, using a robust median + tau * MAD threshold instead of per-head top-k. - Decision Framework: Compare Latent Briefing against prefix caching, summarization, and RAG to pick the right state-transfer mechanism for your stack. - Use Case: In a recursive orchestrator-worker system answering long-document questions, compact the growing trajectory before each worker call to cut worker tokens while preserving answer accuracy. ## Quick Start Ask the assistant to design a KV cache compaction strategy for an orchestrator-worker system where workers need task-relevant slices of the orchestrator trajectory.
