A small simulated sales test found that a multi-agent AI system got past five deliberately triggered stalls after a deterministic reset and fixed reply. A rule-based gatekeeper approved a meeting in all five runs under an exception rule built into the test.
The stall began when a scripted prospect asked the agents to open an NDA link. Their operating rules barred following external links or signing an NDA through chat, so the agents refused but could not move the conversation forward. In the simulation, regular-expression rules checked the pitch and decided whether to approve a meeting; Claude Haiku 4.5 generated the prospect's reply text but did not make that decision.
The recovery step detected the refusal pattern, erased the agents' conversation history and the Director's plan cache, and sent a prewritten message. It left the prospect's memory untouched. The fixed 394-character message cost no model call and met the published exception rule: ask a question about analyst backlog, frame the service as overflow research, and offer an upfront paid proof of concept at the published list price.
Across five runs without recovery, the gatekeeper approved no meetings. With the reset and fixed reply, it approved all five. Both groups used a median of 24 model calls across the full conversation, so the result was not a measured reduction in total calls; the fixed recovery message itself required no new generation call.
A third group asked an LLM to write the recovery reply. An LLM monitor judged those drafts against the agents' operating rules and approved four. A separate, deterministic send check rejected all five under a narrower phrase rule, so the system substituted the fixed text each time. Because that check was stricter than the prospect's acceptance test, the study does not show how the LLM-written replies would have performed on their own.
In an October 8 arXiv preprint, Masaaki Nakatsu and Reno Wang describe this as a known-solution test: the prospect's acceptance rule was published and partly encoded in the agents' instructions and fixed reply. Each arm had five runs in one sales scenario, with one product and one scripted prospect. The result shows the system executing a prewritten route; it does not establish that it can negotiate with real customers or recover generally from agent deadlocks.
Illustrative photo by Shixart1985, licensed under CC BY 2.0. The people pictured did not take part in the study.
