Text Games
Interactive fiction, structured text observations, menu actions, and token-level decisions.
NeurIPS 2026 Competition Track
Foundation Agent Challenge 2026
One agent. A thousand games.
Train a single model and harness to play across text games, web games, and emulated games. The competition runs on Kaggle; this site is the landing page for rules, resources, and updates.
100
public-crawl games
1,000
private-crawl games
40B
parameter limit
The Challenge
Can one agent play many games?
Participants submit one self-contained agent: a model up to 40B parameters plus a harness. The public crawl gives teams a broad training set; the held-out private crawl determines the primary ranking.
Scores are normalized per game, averaged within each runtime category, then averaged across categories with equal weight.
Web Games
Browser games rendered through JS and HTML5, controlled with mouse and keyboard actions.
Emulated Games
Console and handheld titles with video-frame observations and emulator button inputs.
Kaggle Hosted
Kaggle is where the competition happens.
Registration, containers, submissions, and live competition operations will run through Kaggle. Official competition links and materials will be posted here as they become available.
Preview
Built for generalist game-playing agents.
The crawl spans action, platformer, puzzle, RPG, simulation, strategy, adventure, racing, and survival games across multiple runtimes.
Timeline
2026 competition schedule
- Beta launch with public crawl, starter kit, baseline harness, and live leaderboard.
- Hybrid hackathon and official kickoff; rules and crawl freeze.
- Text games sprint.
- Web games sprint.
- Emulated games sprint.
- Final submissions due before private-crawl evaluation and report review.
Rules Snapshot
Single model, offline evaluation, no private-crawl targeting.
Open to everyone, with no team-size or affiliation restrictions.
Submissions must use one model with at most 40B parameters.
External resources required at inference time must ship inside the container.
Commercial or proprietary model API calls are prohibited during evaluation.
Training on, identifying, or reporting private-crawl games is prohibited.
Participants may submit up to three times per day.
Resources
Starter Kit
Kaggle-compatible containers, unified observation and action interfaces, and public-crawl environments.
Baseline Harness
Continual Harness extended across the public crawl, with an open-source student baseline planned for launch.
Participant Support
Tutorials, FAQs, introductory webinars, Discord support, and GitHub Issues for reproducible bug reports.
Organizers
Organizing team
The challenge is organized by researchers across agent learning, game AI, multimodal evaluation, and competition infrastructure.
Seth Karten
Princeton University
Wayne Chi
Carnegie Mellon University
Jake Grigsby
UT Austin
Wenzhe Li
Princeton University
Chengshuai Shi
Princeton University
Alex L. Zhang
MIT CSAIL
Fei Fang
Carnegie Mellon University
Stephanie Milani
NYU / Johns Hopkins
Karthik Narasimhan
Princeton University
Kiran Vodrahalli
Google DeepMind
Yuke Zhu
UT Austin / NVIDIA Research
Chi Jin
Princeton University