← News

Agents & MCPSources: arXiv · 19 Nov 2026

Code-writing agents outperform model-controlled play in real-time game study

A new arXiv paper reports that agents storing learned skills as programs beat a trained neural controller across 170 game tasks, with limits on compute and evidence.

In plain words

Researchers tested two ways for AI agents to play unfamiliar video games where the screen is the only input and the game never pauses. One method has a model write and refine a program between attempts, and the program then plays. That approach made more progress than a model choosing actions moment by moment.

What was announced

The paper, posted to arXiv, frames the problem as self-improvement on an unfamiliar game from a handful of the agent's own attempts, in real-time visual settings where the agent sees only the screen and the game keeps running while the agent thinks. The authors compare two policy types. In the first, a model selects actions during play; the team adapted real-time chunking so committed actions keep executing while the model reasons, and trained a 2B-parameter controller for this interface. In the second, a model writes, tests, and revises a program between attempts, and that program acts during play.

Across 170 GameWorld tasks, with at most eight one-minute executions per run, the code-based approach reached 47.80% mean progress versus 30.90% for the trained controller, and led in all five game categories, according to the authors. Improving the controller's weights on a single game produced only modest gains. The authors conclude the results favor letting a model think offline and storing what it learns in a program that acts in real time. A project repository is linked from the paper.

What it does not show

  • The evidence is the paper's abstract only; the full methodology, task details, and results have not been reviewed here.
  • The reported figures are the authors' own measurements and have not been independently confirmed.
  • The abstract does not describe how the tasks were selected or how general the results are beyond the 170 GameWorld tasks tested.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.