attention-matching

Introduces Latent Briefing, a method for multi-agent systems to communicate via KV cache compaction, achieving 31% fewer tokens with same accuracy and up to 20x speedup.