AI agents manage just 26 percent of orders in Overcooked-inspired test

A new benchmark has put vision-language models in a virtual kitchen, and the results are a useful reality check for anyone treating AI agents as ready-made replacements for human teamwork. In the Overcooked-inspired DuoCook test, the strongest model completed an average of only 26.2 per cent of orders. The other three models scored between 8.8 and 11.2 per cent.¹

The benchmark is built around the sort of problem that makes Overcooked such a frantic co-op game: several jobs need doing at once, the kitchen layout changes what is possible, and a delay in one task can spoil everything that follows. DuoCook recreates that pressure in Unreal Engine, with one vision-language model controlling two virtual players from top-down visual observations.²

What DuoCook actually tests

DuoCook is not a test of whether an AI can recognise an onion or select a button. The model receives a kitchen image, a prop legend, task instructions, live state information and a list of legal operations. It must then assign work to two characters, choose destinations and keep the order pipeline moving while the clock runs down.²

The researchers created 20 maps across five categories. Some are relatively straightforward, while others introduce shared equipment, changing terrain, separated workspaces and hand-off dependencies. That matters because the challenge is not simply completing isolated actions. The agents have to understand which actions are urgent, which resources are blocked, and how one character's route affects the other character's options.

That is where the benchmark becomes more revealing than a tidy software demonstration. A model may know that an ingredient needs chopping, but still send both characters towards the same counter, leave a pan unattended or fail to anticipate that a hand-off will be needed later. In a real kitchen, a human teammate can repair that mistake with a quick shout. An autonomous agent has to notice the problem, update its plan and act before the order expires.

The best model still missed most orders

The strongest evaluated system, identified in the paper as gpt-6-astra, completed 26.2 per cent of orders on average. The remaining models managed between 8.8 and 11.2 per cent. Maps with dynamic obstacles and shared equipment were particularly difficult, as were layouts where ingredients had to pass between separate areas.¹

The researchers also tested a shared notebook that let the agents record observations. Memory was not a universal fix. Claude improved in the notebook condition, while GPT and Gemini performed worse, and API costs increased for all four models.¹ That is a small but important detail: giving an agent more room to remember what happened does not automatically give it better judgement about what should happen next.

The paper measures more than completed orders. It also considers delivery precision, action acceptance and the cost of model calls, which helps separate a system that occasionally gets lucky from one that can reliably coordinate a sequence of tasks. The benchmark's design therefore asks a practical question: can a model keep a changing plan coherent when several goals compete for the same equipment and time?

Why a game-like test matters for AI agents

It would be easy to dismiss the result as a model failing at a colourful cooking game. That would miss the point. DuoCook is a deliberately narrow environment, but its underlying problems appear in many proposed AI-agent applications: dividing work, sharing resources, tracking intermediate goals, responding to changing conditions and knowing when an earlier plan has stopped making sense.²

A coding agent that edits several files, a customer-service system that hands a case between specialist tools, and a household robot moving through a cluttered room all face versions of the same coordination problem. The actions are different, but the need to maintain a useful picture of state is familiar. An agent can be impressive at individual steps and still fail when those steps have to fit together.

The benchmark also makes the limits of broad claims about planning easier to see. A system that completes 26.2 per cent of orders may have learned useful visual and procedural skills, but it is not demonstrating dependable kitchen management. The gap between a successful demonstration and a reliable service is precisely where scheduling, recovery and cost control become important.

The result needs a little caution

DuoCook is a preprint rather than a peer-reviewed paper, so its conclusions should be treated as an early research result rather than a final ranking of today's AI models.¹ The benchmark is also a constructed Unreal Engine environment. It gives researchers control over maps, inputs and scoring, but that control cannot reproduce every ambiguity or physical constraint found outside a simulation.

Even so, the test is valuable because it exposes failure under a repeatable set of pressures. The researchers compare four vision-language models over the same 20 maps and report several measures rather than presenting one cherry-picked success. That makes the result more useful than a short video in which an agent completes one carefully prepared task.

There is another caveat in the name: DuoCook is Overcooked-inspired, not the commercial game itself. The researchers use a cooking environment to study visual planning and coordination; they are not claiming that an AI has mastered Overcooked's full release or can play alongside a human teammate without help.

Our opinion

DuoCook is a better argument for testing AI agents in messy, goal-driven environments than for declaring them ready to run unsupervised. The 26.2 per cent result is not embarrassing because games are trivial; it is interesting because the game-like setting strips away some real-world complexity while preserving the coordination problems that demos tend to hide.

The most telling finding may be the uneven effect of memory. If a shared notebook helps one model and harms two others while adding cost, then the hard part is not merely retaining more information. Agents need to decide which information matters, when a plan has failed and how to negotiate competing tasks. Until systems can do that consistently, a polished single-task demonstration tells us much less than a chaotic kitchen test.

DuoCook does not prove that AI agents cannot coordinate. It shows that coordination remains a demanding capability, and that better-looking interfaces or longer context windows are not substitutes for dependable planning. That is exactly the sort of result the field needs more of.