level: research
when llm agents pass information to each other, the choice of message format matters, but not in a simple way. previous work either shows that structured formats like json cut costs without hurting accuracy, or that imposing structure degrades generation. this study looks at what happens over multiple hops, where copy fidelity becomes key. the authors built a relay testbed where twelve atomic facts are re-encoded across six hops in five formats: free natural language, precision-instructed natural language, json, triples, and key-value pairs. a fixed strong grader scored outputs against programmatic ground truth.
the results show that format effects depend on the relay agent's capability tier. under faithful-relay instructions, a strong relay model was nearly lossless regardless of format. however, weaker models suffered significant degradation when forced to use structured formats, especially over multiple hops. the study also tested a cognitive-load condition and a paired-fork error injection to see how errors propagate. the key finding is that format optimization advice cannot be applied uniformly; it must account for the model's tier.
for practitioners building multi-agent systems, this means that forcing json or other structured messages on weaker models can backfire, causing information loss across hops. strong models can handle structure without issue, but weaker ones need more natural, less constrained formats to maintain fidelity. the research highlights the importance of testing relay chains end-to-end rather than assuming single-hop results will hold. it also suggests that relay instructions and format choices should be tuned per model tier to balance cost and accuracy.
why it matters: designing reliable multi-agent ai systems requires matching message formats to model capability, as forcing structure on weaker models can silently degrade information across relay hops.