Splitting a hard task across several language model agents has become a popular reflex, and the results are oddly inconsistent. Sometimes a crowd of specialists beats one model working alone, sometimes it does worse, and the field has not had a clean account of when to expect which.

The authors offer one, and it turns on how information moves. A single agent keeps its entire reasoning trace in one shared context, so everything it worked out earlier is still sitting there. A multi-agent system splits the work into isolated local contexts joined only by relay messages, and those messages are bounded. Whatever one agent understood has to survive being compressed into a short handoff.

That framing, an information bottleneck, is what makes the inconsistency predictable rather than mysterious. It is the difference between thinking a problem through on one sheet of paper and solving it as a relay where each person passes on a sticky note. Sometimes the note is enough. Sometimes the thing that mattered was exactly what did not fit.