Multi-agent systems using too many tokens? Ramp Labs' new solution cuts consumption by up to 65% without accuracy loss

Multi-agent systems face exponential token growth as reasoning chains extend during task decomposition and repeated model calls—this is a major pain point for large-scale applications. Ramp Labs, an AI infrastructure company, released Latent Briefing: a solution that uses attention mechanisms to identify key context information and directly compresses KV cache at the representation layer, without relying on slow LLM summaries or unstable RAG retrieval.

In LongBench v2 tests, the solution performed excellently: Worker model token consumption was reduced by up to 65%, median token savings for 32k-100k medium-length documents reached 49%, accuracy increased by about 3 percentage points compared to the baseline, and each compression only took 1.7 seconds—20 times faster than the original algorithm. Experiments used Claude Sonnet4 as the orchestrator and Qwen3-14B as the worker model, covering various document scenarios, with optimal compression thresholds varying by task difficulty and document length.