PlayCraft: Designing Engaging Games from Agent Play Experience
Weihao Cheng
cs.AI, cs.CV
2026-11-20
PlayCraft revises 10 games from a vision agent's 10-episode reports. High-intelligence learning gain rises from 12.4 to 34.3 points and leads both baselines at every setting.
Coding agents can already turn a prompt into a browser game that runs. A game that runs is a weak target if nobody wants another round. Design theory treats mastery as a main source of fun: players practice, form strategies, and finish clearly ahead of where they started. That curve usually comes from humans playing and retuning. Generative pipelines rarely optimize it.
Search-based players and controllers trained on one game have been used as stand-ins, with learnability scored as if it were fun. Those players hunt strategies differently from people, and they do not say where they got stuck. Other revision systems do playtest, then chase overall quality. PlayCraft keeps one job. A foundation-model agent that sees only rendered frames plays repeatedly, and a crafter rewrites the game from those reports so practice can raise the score.
Start from a playable game. Alternate play and craft for three rounds.
The player agent sees rendered frames and uses the same keys and mouse a person would. It cannot read the code. Frames arrive every 200 ms. Every 1 second of active time it looks at the last 5 frames and returns 50 control actions of 20 ms each, covering the next second. An episode ends on a clear condition or at 60 seconds. The game pauses while the model thinks, so latency does not spend the budget. A round is 10 episodes. The game resets; the conversation and the written lessons stay. After each episode it records progress, score, decisions, and what it understood, then updates a lesson memo. Action planning uses GPT-6.1 Sol at medium reasoning. Episode review uses extra-high reasoning.
The crafter first estimates perception, control, and strategy the player already had before this game. Ten episodes often never reach expert play, so the 100% reference is extrapolated, not read off the log. It then estimates beginner, intermediate, and expert scores. The expert should finish the objective inside 60 seconds, and that result is 100%. Beginner anchors sit at 10 to 30 percent of that reference, intermediate at 50 to 70 percent. Those strategies should still pay off if the time limit is removed. Gaps are diagnosed across entry difficulty, execution load, available strategies, rewards, pacing, and the performance ceiling.
The design proposal and the code change happen in separate sessions, both on GPT-6.1 Sol: extra-high reasoning for design, maximum for implementation. A change should open a new way to play while keeping the game's defining interaction. The learning value of a rule is weighed against the complexity it adds. Unfamiliar mechanics are not dropped in just to hold beginner scores down. Each change is checked in a real build and in the renderer. The note left for players states only the objective, the rules, and the controls.
Code and Policy share the anchors, the rewrite steps, the model, and the three iterations. Only the evidence changes. Code reads the source and has no player. Policy playtests with 10 controllers that see internal state and carry nothing between episodes, following the generated-policy setup in Recursive Game Creator.
The suite is 10 Three.js games: 2D recreations of Flappy Bird, T-Rex Runner, Breakout, Snake, and Pac-Man, plus prompt-generated 3D games Archery Hunting, Desert Racing, Robotic Arm, Drone Delivery, and Space Shooter. Evaluation runs are fresh. Low is GPT-6.1 Terra with medium planning and extra-high review. Medium is GPT-6.1 Sol with medium planning and extra-high review. High is the same Sol model with extra-high planning and maximum review. Each version gets 3 independent 10-episode runs at each setting.
Score is the episode's final points, except Desert Racing, which uses checkpoints divided by active time. Revision changes the attainable range, so each game and each method is divided by its own maximum among 90 episode outcomes. Learning gain, from Learn2Play Bench, is the mean of episodes 8-10 minus the mean of episodes 1-3, times 100, in percentage points. Learning slope is the least-squares slope of normalized score against episode index, times 100. The macro-average weights the 10 games equally.
Learning gain:
| Setting | Initial | Code | Policy | PlayCraft |
| Low | 11.0 | 7.4 | 12.3 | 13.7 |
| Medium | 17.5 | 23.5 | 16.2 | 31.5 |
| High | 12.4 | 16.1 | 22.3 | 34.3 |
PlayCraft is highest at all three settings. Against the initial games the gaps are +2.7, +14.0, and +21.9 percentage points, with a higher gain in 8/10, 9/10, and 8/10 games. Against Code the gaps are +6.3, +8.0, and +18.2. Against Policy they are +1.4, +15.3, and +11.9.
Learning slope:
| Setting | Initial | Code | Policy | PlayCraft |
| Low | 1.96 | 1.15 | 1.82 | 1.95 |
| Medium | 2.52 | 3.49 | 2.62 | 4.61 |
| High | 2.10 | 2.32 | 3.20 | 4.74 |
At Low, 1.95 is effectively tied with the initial 1.96. The paper puts the deficit at about 0.0033. PlayCraft still beats Code by 0.80 and Policy by 0.13, and is steeper in 7 of 10 games. Medium and High beat the initial games by 2.09 and 2.64.
From Low to High, PlayCraft's gain rises 20.5 percentage points. Code, Policy, and the initial games rise 8.7, 10.0, and 1.4. Slope rises 2.78, against 1.17, 1.38, and 0.14. Most of that jump is Low to Medium: +17.8 gain and +2.66 slope. Medium to High adds only 2.8 and 0.13. Over the first three episodes, High already averages 47.3% of normalized performance and Medium 38.2%. Revision collected its play logs at the Medium configuration, so the stronger player starts higher and has less room left.
On the Medium iteration plot, both metrics finish above the initial game in 9 of 10 titles. Robotic Arm drops hardest after the first revision and stays below its start. At High on Pac-Man, PlayCraft's gain is 3.1, against 22.6 for the initial game and 43.9 for Policy.
Breakout replaces one-hit bricks with underside damage that stacks, a break on the third hit, and side or top hits that clear damaged neighbors. PlayCraft leads both metrics at all three settings. Archery Hunting adds cover and targets that retreat from noise. A quiet hit on a gold weak spot can leave a neighbor in place. It leads at all three settings too. Drone Delivery adds high-altitude wind, shelter only behind buildings, and bonuses for leftover battery and low damage. Medium and High lead. At Low the agent often runs out of battery or crashes before the second delivery, and both metrics finish last. Robotic Arm adds wrist sway under load, randomized can positions, and a rotating deck. Missed grips and blocking remain after placement is learned. At every setting it trails at least one baseline.
The edit is driven by play through the same interface a person has: the screen, the same controls, and lessons carried into the next episode. A small game that already runs can get a learning curve without a human playtest. On the macro-average, both metrics at every intelligence setting are at least as high as Code and Policy. The Low slope is tied with the initial games. The margin grows as the player gets stronger, which fits challenges that pay off for better decisions and cleaner execution.
Two limits are stated in the paper. Agent play and agent learning are assumed to look like human play. Training on human data may pull decisions closer to people, and nobody checked that against human choices, stuck points, or learning curves. Stretching the 60-second budget scales both play time and inference, so longer games stay expensive.
The reports that guide revision come from the Medium setting used at test time: GPT-6.1 Sol, medium planning, extra-high review. High only turns the reasoning up. The games can end up fitted to this model's visual habits. Low swaps in GPT-6.1 Terra and is still not a person. Fun itself was never measured.
Normalization divides by each method's own maximum, so gain is not a reading of difficulty or raw score. In the Medium iteration plot, Robotic Arm finishes below the initial game on both metrics. Three runs of 10 episodes leave visible noise. Pac-Man at High and Robotic Arm show large losses. A macro-average win does not mean all ten games improved. The game pauses during inference, so the agent never faces human real-time pressure. The 10-30% and 50-70% anchors are chosen numbers, and no sensitivity check is reported.