Answers are grounded in this paper and cite their sources.
Agents can collectively hold every fact needed for a decision while no individual has the evidence to justify it. We tested whether naming the location of new activity changes how often language model teams assemble distributed evidence after the primary communication channel closes. Of 163 newly started five-agent teams, 158 qualified and 112 spontaneously wrote to a backup channel. At each first write, we copied the team state into four isolated continuations and randomly assigned the conditions to execution slots. The four other agents received no notice, neutral metadata, a generic activity notice or the same notice naming the channel. The prespecified endpoint was the presence of all three authenticated decision rules in a single actual model input within 60 turns of channel closure. Complete integration occurred in 2, 2, 17 and 112 of the same 112 teams, respectively. Naming the channel increased integration by 84.8 percentage points over generic activity (simultaneous 95% confidence interval 69.1–92.5). All 95 discordant pairs favored naming. Under generic activity, 97 teams read another agent’s task message and 90 received a core rule from another agent. Sixteen submitted a valid solution, compared with 112 under named activity. In this system, linking an event to a channel already visible in the tools changed whether distributed evidence reached one agent and supported a valid decision.
No takes yet. Share an insight, caveat, or question.
Christian Jörg Rustenbach (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: