AI teammates still struggle to cook together: DuoCook tests 20 game kitchens

An AI character that can walk around a game is one thing. Can it actually help a teammate cook dinner, share the same stove and adjust when a kitchen changes? A new experiment suggests that even strong models struggle with this kind of coordination.

Researchers introduced DuoCook on October 22, 2026: a cooking-game benchmark inspired by Overcooked, built in Unreal Engine to test one vision-language AI controlling two characters. The research also offers a practical lesson for creators building games with AI: an impressive demo is not the same as reliable behavior under gameplay pressure.

The researchers built 20 different game kitchens, partly using coding agents

DuoCook contains 20 playable simulation maps in five groups. Some kitchens let both cooks work independently, others force them to share a chopping board or stove, others change passable routes, and others separate the players so they must hand ingredients across.

The environments were constructed with AI coding agents using SimWorld Studio and Unreal Engine. Researchers also performed manual editing to correct layout, visibility, collisions and camera angles, then checked that the required game objects existed. This is AI-assisted level production—not evidence that one prompt created every finished map.

In the test, a single AI model gives instructions to both cooks. The game engine handles movement, while the model decides where to go and which step to perform next. Each action depends on a top-down view of the kitchen, visible equipment, the current recipe, pending orders and game-state data.

The key question: can an AI plan for two characters at once?

Imagine one cook is using the chopping board while soup is already boiling. The better choice may be to send the other cook to prepare a separate ingredient, not queue both at the same place. DuoCook deliberately adds shared equipment, changing floor layouts and handoffs to reveal scheduling failures.

This matters for creators who want NPC coworkers, teammates or assistants rather than characters that only walk on a path. The study tests decision-making and task allocation within its controlled kitchen rules, not natural player conversations or general-purpose game AI.

The results are a warning against assuming an NPC agent 'just works'

The researchers compared four AI models on all 20 maps. Their strongest model, GPT-6 Astra, completed an average of 26.2% of cooking orders, with 64.6% of issued actions accepted. The other three models completed roughly 8.8%–11.2%. Even the best model left most tasks unfinished in this setup. These are the researchers' results, not a GameSlash reproduction.

Complex maps were particularly hard: shared equipment combined with moving terrain, or kitchens where the two cooks cannot cross between areas. A separate experiment gave the agents a shared notebook. That memory sometimes helped but sometimes hurt performance and raised API costs, so adding more memory was not a universal fix.

Each model was run only once per map and condition. The authors warn that the rankings are descriptive and cannot establish how consistently a model will perform over repeated runs. The public manuscript also does not provide episode logs or per-map analysis scripts needed to independently recompute the results.

What does an AI game creator get from this?

The useful pattern is straightforward: create several variations of a task, define valid actions, let the agent observe real game state, then measure how often it finishes jobs—not just whether one recorded demo looks convincing. Also measure the cost of repeated AI calls and how often commands are rejected.

If you are vibe-coding a management, co-op or simulation game, DuoCook offers a research-backed checklist of difficult behaviors to test: competing for one machine, reacting to changing paths, handing items to a teammate and completing tasks on time. It is a benchmark design to learn from, not a plug-and-play NPC system for your own game.

Source: research paper (October 22, 2026) · HTML with original figures · Figures under CC BY 4.0. The study uses an Overcooked-inspired task; it is not an official Overcooked game or a released commercial gameplay feature.