Agentic AI & GenAIConference50min
Agentic Mobile Game Development and Puzzle Solving: Experiments from The New York Times Games Team
Case study of on-device AI game agents for NYT Games: adaptive difficulty, real-time shooter control, and crossword solving using constraint satisfaction. The system runs privately on phones, meeting accessibility rules while fitting tight memory, latency, and battery limits. Tools: Android, JetBrains Koog, and LiteRT.
talk.summaryAiDisclaimer
Shafik QuoraisheeThe New York Times
Games usually handle difficulty with one setting picked once. We tried something different: AI agents that run on the phone during play and adjust the game to the person playing it, moment to moment. It all runs on-device: no cloud calls, no player data leaving the phone.
This is a case study from experimental work on the NYT Games team. We show two agents running live. The first plays an arcade shooter in real time, reading the screen frame by frame, predicting where things are headed, and reacting. The second solves a crossword by searching for a valid fill and backtracking when it hits a dead end. We cover how an agent perceives the board, estimates where the player is looking, and picks settings with constraint satisfaction; WCAG 2.2 accessibility rules are hard constraints it cannot break.
Then the engineering reality of running this inside a game loop on a phone: the agent shares memory with the renderer, thinks in the few milliseconds left between 16 ms frames without dropping one, and stays within a battery budget. We use ARC-AGI-3 to show what these agents cannot do yet
Article in the New York Times about this work: https://www.nytimes.com/2025/12/09/crosswords/why-youre-better-than-a-computer-at-solving-connections.html
Tools used - Android, JetBrains Koog for the agent loop, and LiteRT for on-device inference.
This is a case study from experimental work on the NYT Games team. We show two agents running live. The first plays an arcade shooter in real time, reading the screen frame by frame, predicting where things are headed, and reacting. The second solves a crossword by searching for a valid fill and backtracking when it hits a dead end. We cover how an agent perceives the board, estimates where the player is looking, and picks settings with constraint satisfaction; WCAG 2.2 accessibility rules are hard constraints it cannot break.
Then the engineering reality of running this inside a game loop on a phone: the agent shares memory with the renderer, thinks in the few milliseconds left between 16 ms frames without dropping one, and stays within a battery budget. We use ARC-AGI-3 to show what these agents cannot do yet
Article in the New York Times about this work: https://www.nytimes.com/2025/12/09/crosswords/why-youre-better-than-a-computer-at-solving-connections.html
Tools used - Android, JetBrains Koog for the agent loop, and LiteRT for on-device inference.
Shafik Quoraishee
Shafik Quoraishee is a machine learning engineer, game developer, and technical writer currently working as a Staff Games ML Engineer at The New York Times. His work there focuses on AI and tooling for games including Wordle, Connections, Strands, and Crosswords. He has contributed to internal puzzle analysis systems, multiplayer game features, and experimental ML systems tied to gameplay and accessibility.
Before joining the Times, he worked across media and entertainment companies including the National Basketball Association, Business Insider, and MTV / Viacom, where his engineering work centered on mobile experiences, game systems, and machine learning applications in consumer products.
Earlier in his career, he also worked as a research engineer with the U.S. Department of Defense on intelligent sensor systems and related military research projects.
Outside of industry roles, he has published technical writing and research-oriented articles on topics such as handwriting recognition, neural networks, visual question answering, facial recognition, computational modeling, and AI reasoning. He has spoken at developer and AI conferences, including discussions about how AI systems solve games like Connections and what those limitations reveal about reasoning and cognition.
His broader interests appear to span machine learning, computational biology, physics, simulation systems, and generative AI.
Before joining the Times, he worked across media and entertainment companies including the National Basketball Association, Business Insider, and MTV / Viacom, where his engineering work centered on mobile experiences, game systems, and machine learning applications in consumer products.
Earlier in his career, he also worked as a research engineer with the U.S. Department of Defense on intelligent sensor systems and related military research projects.
Outside of industry roles, he has published technical writing and research-oriented articles on topics such as handwriting recognition, neural networks, visual question answering, facial recognition, computational modeling, and AI reasoning. He has spoken at developer and AI conferences, including discussions about how AI systems solve games like Connections and what those limitations reveal about reasoning and cognition.
His broader interests appear to span machine learning, computational biology, physics, simulation systems, and generative AI.