Agent Based Playability Verification for Automated Game Design

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 233 papers48 conceptsfrozen 2026-10-04

In most creative domains, "quality" is subjective. A painting is beautiful if it moves the viewer; a story is good if it captures the imagination. But in video games, there is a brutal, binary prerequisite for quality: functionality. A level can be visually stunning, but if a player cannot reach the exit, it is a failure.

Julian Togelius, Antonios Liapis, and Georgios N. Yannakakis argue in Computational Game Creativity that games are the "killer app" for AI creativity precisely because they demand this orchestration of aesthetics and hard functional constraints. To solve this, researchers have moved toward a "generate-and-test" loop where automated agents act as the ultimate critics, ensuring that what an AI creates is actually playable.

The Playability Gap

Traditional Procedural Content Generation (PCG) often relies on rules or patterns. Modern approaches use Generative Adversarial Networks (GANs) or Large Language Models (LLMs) to "hallucinate" game content. However, as Xinyu Mao and colleagues note in Procedural Content Generation via Generative Artificial Intelligence, these models lack an inherent understanding of game logic. A GAN might generate a perfect-looking Super Mario Bros. level that contains a jump one pixel too wide for Mario to clear.

This creates the "Playability Gap." To bridge it, researchers integrate automated agents—ranging from simple pathfinders to complex planning systems—directly into the generative pipeline.

Agents as Functional Critics

The most straightforward way to verify playability is to have an AI "play" the level before a human ever sees it. In Evolving Playable Content for Cut the Rope through a Simulation-Based Approach, Noor Shaker and her team used a physics-aware agent to test levels for the puzzle game Cut the Rope. If the agent can reach the goal, the level is valid.

However, this introduces the negative proof problem: if the agent fails, is the level unplayable, or is the agent just not smart enough? Shaker highlights that the bottleneck in automated design isn't generating the content, but the computational cost of evaluating it with "good enough" agents that don't discard high-quality, difficult levels just because they lack the precision to solve them.

Steering the Latent Space

Rather than just filtering out bad levels at the end, researchers now use agents to "steer" the generative process. Vanessa Volz and Jacob Schrum demonstrated this by combining GANs with evolutionary search in Evolving mario levels in the latent space of a deep convolutional generative adversarial network.

By using an agent’s success rate as a "fitness function," they can navigate the GAN’s latent space—the mathematical "map" of all possible levels the GAN can make—to find regions that are both visually authentic and functionally sound. Similarly, Matthew Christopher Fontaine uses Latent Space Illumination to map out a diverse palette of levels, ensuring every point on that map is verified by an automated pathfinder.

Beyond Levels: Verifying Mechanics and Logic

Playability verification isn't limited to 2D layouts. In Automatic Game Design via Mechanic Generation, Alex Zook and Mark O. Riedl use automated planning agents to verify the actual rules of a game. If the system generates a new mechanic (e.g., "jumping on an enemy grants a double-jump"), the planner attempts to use that mechanic to reach a goal state. If the goal is unreachable, the mechanic is discarded as broken.

This logic extends to 3D environments. The LUMINOUS framework by Yizhou Zhao and colleagues generates interactive household scenes for Embodied AI. They use an automated planner to execute multi-step tasks (like "find the apple and put it in the fridge") to prove that the procedurally generated 3D house is physically navigable and interactively consistent.

Mixed-Initiative Design and Real-Time Feedback

The ultimate goal of this research is often to assist, rather than replace, human designers. Tools like RL Brush, developed by Omar Delarosa and Ahmed Khalifa, use Reinforcement Learning agents to provide real-time "nudges" to designers Mixed-Initiative Level Design with RL Brush.

As a designer draws a level, the AI acts as a continuous feasibility checker. In Matthew Guzdial’s work on Morai Maker, the system uses A* search and survival analysis to tell the designer instantly if a section is playable. This allows creators to build levels that exceed their own physical playing ability, relying on the AI to confirm that a high-difficulty jump is technically possible.

From Reachability to Experience

The field is now moving from simple "can the agent win?" checks to "how does the agent feel?" Julian Togelius and colleagues explore this in MiniDungeons 2 by creating "procedural personas"—AI models that mimic specific human playstyles (e.g., a "runner" who ignores treasure vs. a "collector" who explores every corner).

By testing a generated level against a population of these personas, designers can verify not just if a level is solvable, but if it is fun, balanced, and appropriately challenging for different types of players. As Swen E. Gaudl demonstrates with Fluidic Games, high-speed agent simulations can calibrate difficulty on the fly, ensuring that the transition from player to designer is seamless and supported by a robust functional safety net.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 18 papers
  2. Version 1 · 2026-10-04 · 33 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_agent_based_fitness_evaluation_c70277ad,
  title = {agent-based fitness evaluation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/agent-based-fitness-evaluation/rdr_c70277ad9a2c48b6a97ce877ca17bc3a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_agent_based_playability_verification_for_c70277ad_v1,
  title = {Agent Based Playability Verification for Automated Game Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/agent-based-playability-verification-for-automated-game-design/rdr_c70277ad9a2c48b6a97ce877ca17bc3a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}