SEE THE GAME
Read the game brief and track the action through the current screen and recent frames. Motion and changing state matter, not just a single screenshot.
GENERALIST GAMING AGENTS / 2026
Can one agent pick up an unfamiliar real-time game, figure out the rules, and act at the right moment? We're building an open benchmark to find out.
INDEPENDENT RESEARCH INITIATIVE · PUBLIC COMPETITION IN PREPARATION
01 / THE MISSION
Many game-playing agents are specialists. We test whether one system can adapt across unfamiliar real-time games, where timing matters as much as action selection.
A generalist must perceive a new interface, learn its rules, plan ahead, and time its actions as game state evolves, without custom code for each game. Real-time play brings perception, reasoning, and control together. Those same capabilities matter for computer-use agents and embodied systems.
SEE HOW WE EVALUATE02 / THE BENCHMARK
One offline model and one harness must handle the rules and timing of 32 real-time games. Four sample games support development; the other 28 are withheld.
Read the game brief and track the action through the current screen and recent frames. Motion and changing state matter, not just a single screenshot.
Choose what to press, when to press it, and how long to hold it. Every game uses the same mouse-and-keyboard interface, with no game-specific adapter.
Scores come directly from game state, normalized to 0–1—no subjective model judge.
One model + harness
16 public / 16 private leaderboard games
Offline, reproducible evaluation
Open research resources
03 / THE COMPETITION
Build one offline agent that plays 32 unfamiliar real-time browser games. The upcoming Kaggle competition has a $50,000 prize pool. Test adaptation, planning, and timing-sensitive control with one model and harness.
04 / RESEARCH DESIGN
Real-time games bring perception, planning, exploration, and control into one setting. An agent must choose not only what to do, but when and for how long. The goal is reproducible evidence of transfer across environments.
Four sample games are provided for development. The remaining 28 are withheld. Sixteen games contribute to the public leaderboard and sixteen to the private leaderboard.
The right action at the wrong moment can fail. Agents must track changing game state and coordinate when to press, hold, release, or switch inputs.
Each game exposes a programmatic performance signal instead of relying on screenshot interpretation or an AI judge.
The organizers are preparing shared harnesses, reference baselines, and pinned evaluation assets.
A larger organizer-operated game collection is planned for subsequent studies of cross-game transfer.
05 / THE PEOPLE
An independent, collaborative research effort. Organizers are listed in citation order.

University of Hamburg
Princeton University (visiting)

ETH Zürich
Princeton University (visiting)

University of Hamburg
Princeton University (visiting)

Carnegie Mellon University

Microsoft Research New York
New York University (visiting)

Princeton University

Carnegie Mellon University

Princeton University

Princeton University

Google DeepMind

Princeton University
06 / FAQ
Yes. The central challenge is a single offline model-and-harness pair that handles the rules and timing of all 32 real-time games, rather than a different agent for each title.
Game state changes over time. Agents need to track motion and choose when to press, hold, or release controls, not just select an action from a static screen. Real-time describes the gameplay; the final competition rules will specify evaluation timing and runtime limits.
It gets the game brief, objective, controls, and visual observations. It interacts through a universal mouse-and-keyboard interface.
Four sample games are provided; the other 28 games are withheld during the competition. “Public” and “private” refer to the two 16-game leaderboard splits, not which game builds are distributed. Final rules and submission instructions will be available on Kaggle at launch.
The competition is in preparation. Launch timing, submission instructions, and final rules will be announced on the Kaggle competition page.
All 32 games have MIT-licensed code. We have requested MIT licensing for original competition resources and data, and for winning submissions’ original code. Third-party components, assets, and pretrained weights retain their applicable licenses. The final Kaggle rules will govern competition participation.
THE QUESTION REMAINS
Help test generalization where perception, planning, and real-time control meet.
FIND US ON KAGGLE