AI agents checked out before reading the messages that mattered

A group of AI assistants can place an order before the assistant checking out has read an important request already sent to it. In research published September 30, Sahan Paliskara, Nattaput Namchittai and Andrew Lampinen found that a single coordinator honored a targeted user's request about twice as often as a team of personal agents in simulated group orders and bookings.

The MAMUBench study examines a familiar coordination problem in a new setting: each assistant serves a different person, but they share a cart, budget or calendar. The research covers 77 scenarios across four environments and several models. Its 25 personal-assistant scenarios include requests for side dishes instead of an entree, offers to pay extra for other travelers, and safety-related constraints. The orders and reservations are simulated.

Simply giving the agents a communication channel did not solve the problem. Among team episodes that committed an order or booking, the targeted request had been sent to the committing agent in 94–99% of cases. Yet it was in that agent's context at checkout in only 47–67%. Reading it was not sufficient either. Across the tested models, 65–77% of team episodes included at least one edit that violated the request.

Putting a team leader in charge of checkout helped about as much as simply rerunning failed episodes. A software guard worked better: it blocked checkout while the committing agent still had unread messages. When previously failed episodes were rerun with this guard, 73.1% were recovered for Claude Opus 5 and 62.3% for Claude Sonnet 5. Those are recovery rates within the failed group, not overall success rates for all orders.

The agents also made unsupported claims about other people, including supposed agreements. Claude models judged fabrication and consent, and those judgments were not validated against human reviewers. The analysis does not establish that any one observed behavior caused the entire performance gap. Real users, live services and newer models require separate tests.

The practical challenge is concrete: an assistant arranging dinner must carry everyone's constraints through to the final action. Checking for unread messages before checkout deserves testing alongside adding more agents. Users should still verify critical conditions, including allergies, timing and cost sharing, on the final confirmation screen.

Independently adapted from research by Sahan Paliskara, Nattaput Namchittai and Andrew Lampinen, with technical detail condensed. The original paper is licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/.

Original source published: arXiv, October 1, 2026, 02:49:25 Beijing time / September 30, 2026, 2:49:25 PM EDT.