Diffusion language models generate answers through a series of revisable continuous states, yet multi‑agent systems usually communicate only after decoding text. We ask whether a single diffusion trajectory can embed a compact hidden message into another and whether the receiver actually uses its meaning.\ In the Hidden Messages experiment, the sender compresses its denoising state into four signed dense vectors; the receiver reads them while preserving its own context, canvas, and trajectory. Two interventions probe the channel: removing the message tests if the channel matters, and substituting it with a vector from another example tests if the content matters.\ Across controlled tasks, deranging the message eliminates a +63.6‑point advantage on Tiny‑A2D and a +100.0‑point advantage on Dream‑7B, while targeted substitutions steer the receiver toward the substituted fact.\ On QASC, where two supporting facts are split across agents, matched communication reaches 80.8% accuracy, but a wrong‑example message drops it to 46.8%, a content effect of +34.0 points.\ The surprising finding is temporal: a single late transfer attains 87.8% accuracy, surpassing four recurrent exchanges by +7.0 points. These results provide causal evidence for continuous hidden communication between diffusion‑LM agents and show that when information is transferred can matter more than how many exchanges occur.\ \ Review