A new paper on arXiv introduces Thought Blocks, a method for multi-agent systems that aims to reduce the overhead of inter-agent communication. The authors note that LLM-based agents typically exchange text, which forces each agent to repeatedly generate and re-encode messages. This is computationally expensive, especially as the number of interactions grows.

The paper explores latent communication as an alternative, where agents share internal representations instead of words. However, the abstract points out that even latent messages can be costly to construct. Thought Blocks address this by compiling agent trajectories—sequences of reasoning steps—into compact states. This allows communication to happen non-autoregressively, potentially avoiding the step-by-step generation bottleneck.

The provided abstract does not include experimental results, so the practical gains remain unspecified in this excerpt. Still, the approach offers a concrete direction for reducing communication overhead in multi-agent LLM systems, a key bottleneck as such systems scale.