Multi-agent systems built on large language models (LLMs) predominantly communicatethrough natural language text. We demonstrate that this text-based communication acts asa severe information bottleneck, analogous to the children’s game of Chinese Whispers, whereeach retransmission introduces compounding distortion. We propose Blindsight Transport, amethod that replaces text-based inter-agent communication with direct hidden state transferthrough the transformer’s residual stream. By splitting a transformer into an early-layercourier agent (layers 0–1) and a late-layer receiver agent (layers 2–N), information flows asraw activations rather than generated text. We prove that this yields KL divergence of exactlyzero from baseline—a lossless channel—whereas standard text-based handoff produces KLdivergences of 2.4–9.3 across all test conditions. We present seven experiments using GPT-2family models (82M–355M parameters): (1) a Chinese Whispers comparison showing 8/8test wins for hidden state transport, (2) multi-agent chains of 2–10 agents where text-basedcommunication causes immediate catastrophic signal destruction while hidden state transportmaintains KL = 0 regardless of chain length, (3) lossless serialization of hidden states to disk,(4) scaling verification across model sizes, (5) cross-model transfer between architecturallydistinct models achieving 29–46× improvement over text, (6) cross-dimensional transfer vialearned linear projection achieving 75× improvement over text, and (7) qualitative analys