Event-triggered activity notifications improve collective information recovery after communication disruption in large language model teams

Abstract Background. A team can retain an alternative communication channel after its primary channel fails, yet evidence posted there may never reach a teammate. This study tested whether notifying agents of activity on the alternative improves collective information recovery. Methods. Eighty qualified five-agent teams each supplied four isolated branches. In the randomized paired comparison, both branches lost the primary channel and retained identical rights to read and write on the alternative. One branch relied on agent-initiated checks (B0). In the other (B+), the first successful write also triggered an activity notification carrying no task evidence for each teammate. The primary endpoint required all three authenticated task-critical sources together in at least one submitted model input by cycle 12. Results. The endpoint occurred in 2/80 teams (2.50%) under B0 and 61/80 (76.25%) under B+. All 59 discordant pairs favored B+. The paired risk difference was + 73.75 percentage points (conservative ≥ 95% confidence interval + 55.82 to + 84.05, exact two-sided McNemar p = 3.46945 × 10⁻¹⁸). First writes occurred in 60 and 61 teams, respectively. A teammate’s authenticated core source reached a model input in 20 and 61 teams. Conclusions. Enabling the conditional notification rule increased complete information recovery in this task family and model configuration. The arms diverged mainly after the first write. The observed stages leave the contributions of attention, channel discovery and timing unresolved.