Playful Intelligence: Bringing Generative Behavior to Games

Unlocking a new generation of adaptive and intelligent games with generative behavior

Games and Artificial Intelligence (AI) have been connected in a virtuous cycle since the early days of computing. Games, with their clear rules and defined scope, have offered both a proving ground and a playground for AI, continuously stretching its definition and pushing the boundaries of what's possible. In turn, that innovation has fed back into the games industry where AI techniques have become part of the game design toolkit for everything from creating environments and storylines, to advancing game mechanics.

At the heart of this connectedness are AI agents. These entities, central to nearly all of AI’s breakthroughs in games, are defined by their ability to take action to achieve their goals. For decades, agents have been created to compete against humans, mastering increasingly complex games as technology has advanced.

The earliest successes for AI agents in games against human professionals happened in checkers and chess in the 90s, when the checkers agent Chinook and chess agent Deep Blue defeated human world champions in their respective games. These victories in what researchers call “deterministic perfect information” games, where there are no random events and the entire game state is visible, were achieved primarily through search techniques. The agents looked ahead through long sequences of future moves to plot a likely course to victory.

As AI research advanced, games with random chance and hidden information became the next frontier. TD-Gammon surpassed human professionals at backgammon, Polaris was the first agent to beat human poker pros in a meaningful match, and Cepheus solved heads-up limit Texas hold’em poker. Instead of search, these agents were powered by learning: using experience from previous games to improve their performance over time, until they were able to surpass human professionals.

However, while the breadth of games addressable had increased, the agents themselves were still very limited. TD-Gammon could only play backgammon, and Polaris and Cepheus could only play poker. The underlying algorithms were general, but a large engineering effort was required to tackle each new game.

The deep learning revolution in the 2010s started to change this. In 2015, DeepMind's DQN agent learned to play a variety of Atari 2600 games, surpassing human performance on many. This was a step-change moment and demonstrated a move from manual implementation being required for each game to agents that could adapt themselves to new games. Following this, agents quickly mastered games like Go (AlphaGo 2016, AlphaZero 2017), no-limit Texas hold’em poker (DeepStack 2017, Libratus 2017, Pluribus 2019), Dota 2 (OpenAI Five 2017), Quake 3 Capture the Flag (FTW 2019), StarCraft (AlphaStar 2019), and Diplomacy (Cicero 2022).

These more flexible agents not only learned from experience but also learned how to perceive and interpret inputs and choose appropriate behaviors. They were more easily adaptable to complicated games that had graphical inputs instead of discrete states. However, several drawbacks remained: they required even more training time than earlier agents, were still specialized for specific games, and could not adapt well to environmental changes or human behavior. Sometimes, a human could beat them by playing irrationally, to take the agent out of its training distribution. And critically, these powerful agents often made gameplay joyless for human players. Losing to a superhuman or optimal opponent, without understanding why, is simply not fun.

With the emergence of foundation models over the last few years, and in particular their combination with Reinforcement Learning, the scope of possibilities for AI and games has increased dramatically. While agents that could search future states and learn from experience were the key components in the past, the new ability to easily incorporate prior knowledge and communicate with natural language has opened the door to creating agents and game experiences that have never before been possible.

Today, agents enabled by generative AI can execute a deeper, more human, more dynamic kind of intelligence in gaming — a playful intelligence that is less about achieving superhuman feats and more about delivering human-centric experiences. Game characters and systems can now be aware of their role in the narrative to support the desired player experience, instead of only playing to win. General out-of-the-box competence across games is now possible, instead of mastery of only one particular game at a time.

This latest evolution represents a huge opportunity not just for AI development but also for the entire gaming industry.

Behavior is all you need

Delivering on this massive opportunity is the primary reason we started Artificial Agency.

We’ve spent our careers researching AI agents, but generative AI has opened up a once in a lifetime opportunity to lift our innovations out of the lab and deliver them to the industry. Though there’s been a lot of interest and energy around introducing generative AI powered agents into gaming, the industry, so far, has primarily focused on enhancing the conversational capabilities of non-playable characters without fully exploring the broader possibilities. We see this as an extraordinary underutilization of the technology's potential.

Language models have been transformational because they enable agents to have natural conversations — until now, a uniquely human ability. However, there are a vast array of more primitive and essential actions an AI agent can perform that can be especially engaging and meaningful in a game environment.

The magic of AI agents is that they can tell us what they want to do and how they want to behave in an environment. So why not empower them fully to do it? This fundamental concept fuels our focus on what we call generative behavior. For us, this is where the true magic of agents lies and the potential to create dynamic, interactive systems that enhance the gaming experience is immense.

Level One

Today, we’re thrilled to be launching Artificial Agency from stealth and announcing $16M (USD) in funding from Radical Ventures, Toyota Ventures, Flying Fish, Kaya, BDC Deep Tech, TIRTA Ventures, and others to power this vision and accelerate the development of our flagship product — an AI-powered behavior engine that enables game developers to embed runtime decision-making seamlessly into any aspect of a game, delivering a gaming experience that feels truly alive.

Our technology, which we are developing in close collaboration with several notable AAA studios, empowers game developers to transform characters and other decision-making systems in games into individualized AI agents with perceptions, actions, personalities, and goals. The result is a more immersive and entertaining experience where the AI keeps the player engaged in a game that is expressive, creative, and delivers the player experience that game designers have always envisioned.

For us, Artificial Agency is the culmination of our lifelong experiences as academic and industrial AI researchers, our careers in AAA game development, and our love of gaming as a pastime and as an art form. We’ve built a team around us that includes some of the best gaming talent in the world, people who know exactly what it takes to build a world-changing game and who are passionate about unlocking creative superpowers for studios of all sizes.

We’re right at the beginning of our journey, only on level one, but so excited for what’s ahead. We have the team, the technology, the funding and the partners to deliver this next generation of agents and make entirely new types of gameplay possible.