Agentic PCG: Procedural Content Generation via Tool-using LLMs

Framework

Perceive, Reason, Plan, and Edit

Instead of asking the LLM to directly generate an entire game level in one shot, we wrap the game as an interactive environment, similar to an RL environment. This allows the environment to handle gameplay dynamics, evaluate the current level, and provide structured feedback based on level metrics. Within this loop, the LLM agent can perceive the current level state, reason about what should be improved, make editing plans, and iteratively modify the level.

Binary Door

Lode Runner

Zelda

Sokoban

Our framework supports both static and dynamic game environments. In some tasks, the level can be evaluated directly from its structure using metrics such as tile counts, connectivity, or solvability. In more dynamic environments, evaluation can also depend on simulated gameplay, such as the actions taken by a deterministic A* agent interacting with the level. This makes it possible to provide rich environment feedback ranging from simple structural statistics to behavior-based signals.

SMB Simulation Run

SMB Quality Optimization

See detailed optimization process below ↓

Tools

Tool-Using Level Design

The tool set can include both simple and complex operations. At the simplest level, the agent may place individual tiles, draw lines, or edit patches of the map. But the framework is not restricted to these primitive edits. In the Binary Maze domain, for example, the agent can also invoke classic PCG algorithms that are naturally suited to maze generation, such as binary space partitioning and tree-search-based diggers.

Step 1 — generate_digger

Rationale
The current level is completely solid, resulting in zero connected regions and zero path length. To begin, I need to generate a traversable space. The generate_digger tool is an excellent choice for this initial step because it guarantees a single connected region. By setting a relatively low stop_size, I aim to create a sparse, tunnel-like structure, which tends to produce longer shortest paths.

Plan
1. Use generate_digger to create a new level with a single connected region.
2. Set stop_size to a low value (0.2) to encourage generation of a long path.

Step 5 — place_tile (walls)

Rationale
Previous attempts to use generate_bsp were rejected without changes. Switching to a manual editing strategy. The current path length of 116 is far below the target of ~256 because the level contains many long, straight corridors and open areas that create shortcuts. Strategically placing walls will block major thoroughfares, forcing movement into more circuitous routes.

Plan
1. Place a vertical wall to bisect the long horizontal corridors on the right side.
2. Place a horizontal wall to separate the large open areas in the bottom half.
3. These walls will create significant detours to increase the longest shortest path.

Step 25 — generate_bsp

Rationale
Manual edits repeatedly failed by either disconnecting the level or not improving the path length. The generate_digger attempt also produced a much shorter path. Trying generate_bsp instead — unlike organic caves from digger, BSP creates a structured, maze-like layout of rooms and corridors inherently better suited for creating long paths.

Plan
1. Acknowledge the failure of both incremental edits and generate_digger.
2. Use generate_bsp with splits=5 to generate a complex map with many small rooms.
3. Evaluate the new metrics.

Step 39 — reroute corridor

Rationale
Returning to a robust strategy of rerouting a major corridor. Moving the connection between Chamber 2 and Chamber 3 from the far right side (x=55) to the far left side (x=10). This will force any path traversing these chambers to travel their full width, adding substantial length to the overall path.

Plan
1. Close the existing passages between Chamber 2 and Chamber 3 by filling with walls.
2. Create a new single passage on the far left at x=10.
3. This rerouting forces the main path to traverse almost the full width of both chambers.

Step 40 — fine-tune to target

Rationale
The previous edit was very successful, but it overshot the target of ~256, landing at 285. To reduce the path length back towards the target, adding a shortcut: a new passage between the same two chambers at a more central location (x=25). This should reduce the path length by approximately 30, bringing it to around 255.

Plan
1. Create a new vertical passage through the wall at column x=25.
2. Place empty tiles at (y=13, x=25) and (y=14, x=25) to form the shortcut.
3. This provides a shorter route than the existing one at x=10, reducing the path closer to 256.

Binary Maze optimization: from initial digger generation to BSP, manual corridor rerouting, and precise fine-tuning to hit the target path length.

Language Control

Free-Form Language Instructions

Beyond functional constraints, the framework can also incorporate free-form natural language instructions. This enables open-ended language control on top of explicit metric-based requirements, so the agent can satisfy concrete functional targets while also adapting to higher-level design goals such as theme, story, or intended player experience.

Creative Level 1

Creative Level 2

Mario levels generated with natural language instructions guiding theme and structure alongside playability constraints.

Controllability

Targeting Different Controllable Metrics

The agent can be directed to optimize toward specific target values for different controllable metrics in each game domain. Each grid below shows multiple trials where the agent targets a different metric value per trial, demonstrating fine-grained controllability over level properties.

Binary Maze Show Final Maps Binary Maze final maps