Skip to content

Does Your Choice Survive an AI Group Order?

In simulated group orders and bookings, one AI assistant with access to everyone's requests and notes was better at keeping a person's explicit choice than separate assistants that had to share what they knew.

In one simulated dinner, 4 colleagues share a $92 food budget. One diner has a favorite certified gluten-free meal. But his written instructions say not to spend the team’s money on it. He wants 2 gluten-free side dishes instead, leaving more money to feed the others.

Ordering his favorite meal would miss the point. He has already decided how to compromise. The assistant’s job is to keep that decision intact while helping put together dinner for everyone.

A September 30 preprint, Worse Together, asks whether separate AI assistants can do that. Researchers constructed 25 scenarios involving food orders, restaurant reservations, and trips. In each, they checked whether the final order or booking honored one person’s explicit request.

One assistant serving everyone did substantially better than a team of separate assistants that could message each other. With Claude Opus 5, the rounded rates were 88% for the single coordinator and 44% for the team. The coordinator also did better with the other 3 models tested in these scenarios.

There was an important difference in what they knew. The coordinator received everyone’s opening requests and could read everyone’s relevant notes. Each separate assistant started with its own person’s request and notes. To understand the group, it had to learn from the others.

This is a comparison of those whole arrangements. It does not isolate the effect of adding more agents. Nor do the percentages measure everyone’s satisfaction: they measure whether the designated request survived in the final shared result. Placing no order counted as failing to fulfill it.

Choosing dinner is only the beginning

The real task already separates choosing from ordering. In DoorDash’s documented business workflow, the creator shares a link and participants add their own items, with a display of their remaining budget. The creator then checks the delivery and payment details and places the order.

Several people contribute, but one person takes the final step. If assistants handle those contributions, the assistant placing the order needs to understand more than its own person’s dinner.

The diner choosing sides makes this unusually clear. His notes describe both a favorite meal and a standing instruction for tight group budgets. The cheaper dinner is his choice, not a sacrifice an assistant has decided to impose. Knowing his dietary restriction alone would not tell the assistant which gluten-free dinner to order.

The researchers specify a cart that honors his note and fits just under the shared cap. They do not report a real dinner or an observed sequence in which agents changed that cart. The authored setup shows what the test is asking assistants to preserve: a person’s decision about how their meal should fit with everyone else’s.

A sent message can still arrive too late

The simulation already restricts food-cart checkout to its creator. Giving one assistant the final button does not guarantee that it knows the others’ wishes.

The researchers found orders and bookings finalized before the relevant request reached that assistant’s conversation, even when another assistant had already sent it. The request was kept more often when it was available to the assistant making the final decision.

The timing helps explain how this can happen. The separate assistants work at the same time. A message starts the recipient’s next turn after its current turn ends; it does not interrupt an assistant already taking actions. The test stops at the first confirmed order or booking. A request can therefore be on its way while the assistant places an order without it.

For the diner choosing sides, telling his own assistant is only the first step. That assistant must convey the choice to whoever places the order, and the information must reach them while the cart can still change.

The researchers tested a rule at that point. They selected previously failed Claude team attempts and reran them unchanged or with confirmation blocked while the finalizing assistant had unread messages.

For both Claude models, the rule kept the request more often than an unchanged rerun. This comparison counted only reruns that confirmed an order or booking and passed the scenario’s validity checks. It tests recovery from selected failures, rather than success on a fresh set of orders.

Assigning confirmation to a designated lead produced recovery rates close to the unchanged reruns. In this test, waiting for messages made a bigger difference than assigning a particular assistant the final step.

But message timing is not the only difference worth considering. In a smaller test on selected failures, giving every assistant everyone’s notes also improved the results. The coordinator begins with access to those choices together. Separate representatives have extra work to do before they can see the same picture.

What the finished order leaves out

The simulation does not reproduce all the real platform’s rules. DoorDash’s business guide describes per-person spending limits, while the simulated dinner has a pooled cap. The guide says the creator can remove items but cannot add them for participants. The simulation allows broader editing of other people’s items. Those differences matter to how a choice can be changed.

The study therefore supplies no failure rate for actual DoorDash orders. It is a new preprint awaiting independent testing, using scenarios where people must share a limited budget or booking. Its message rule is a useful result within that setup, not a demonstrated fix for every shared decision.

The better-performing coordinator also leaves a question open. It can read everyone’s requests and notes. Its advantage does not tell us how much personal information people should have to share to make one order work.

Still, the test asks a more demanding question than whether dinner got ordered. A cart can fit the budget and be placed by the right assistant while losing the choice it was supposed to honor. In the constructed dinner, the diner had asked for 2 sides so the others could have more. Preserving that request was part of ordering his meal.

Sources

Understand How People and Agents Work Together.

Practical essays, experiments, and lessons from building software for people and agents.

Occasional emails, sent when there is something worth reading. Free to subscribe. Unsubscribe at any time.