Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
When relaying information through multiple LLM agents, the format of the message affects fidelity differently based on the relay's capability tier; a weak 1.5B relay shows an 8.7-fold increase in format-dependent recall variation over six hops, whereas a strong relay remains nearly lossless. Structure, like JSON, provides a faithful channel that localizes errors but doesn't correct them, making format choice dependent on the weakest relay in the pipeline. This changes how one should optimize message formats for production LLM agent chains.
A weak 1.5B LLM relay widens the six-hop recall gap between message formats by 8.7× (2.3 to 20.5 points), flipping which formats survive multi-hop pipelines. If your production agents include any low-tier relays, JSON’s fixed-key schema resists drift best, while rigid formats like triples or key-value pay an encoding toll—so format choice must match the weakest link or risk silent fact decay.