latent-briefing

What problem does it solve?

Replaying the full orchestrator reasoning trajectory to every worker inflates token usage and introduces latency, while summarization loses critical context needed for downstream tasks.

Core Features & Use Cases

• Representation‑level sharing: Retains only the KV positions most relevant to the current worker task, avoiding full‑text handoff.
• Task‑guided attention scoring: Uses the worker’s task prompt to score trajectory tokens and build a shared global mask.
• Robust MAD thresholding: Applies median‑plus‑MAD to decide which KV entries to keep, adaptable to different workloads.
• Use Cases: Hierarchical multi‑agent pipelines, recursive language model orchestration, and any system where workers require selective slices of the orchestrator’s latent state without costly text replay.

Quick Start

Ask the agent to compact the orchestrator’s KV cache for the upcoming worker task.