Game-Creation-Agents (GCA) turns a natural-language request into a runnable Phaser 4 browser game. The pipeline first builds a multimodal game design document that couples mechanics and scene contracts to grounded visual references. It then projects that document into a runtime scaffold, machine-readable contracts, and deterministic tests. Eight specialized agents next carry out production: scenes are implemented serially, assets are produced in parallel, and one reduced manifest joins the two branches into a single build. Build-linked real-UI checks and a code-blind critic then verify that build and decide whether the game may ship. Each blocker returns to the agent that owns the affected artifact for a bounded repair, and the game is released once all the gates pass. We evaluate finished games separately, through their rendered interfaces alone, over prompt-matched tasks run under a common protocol. GCA achieves stronger requirement satisfaction, core mechanics, content depth, and functional visuals than a comparable open pipeline. In a side-by-side human study, raters also score GCA's games higher on every dimension and prefer them overall.
Figures & tables
System
Primary scope
Executed or rendered evidence
Role of the outcome
GameGPT ( Chen et al., 2023 )
Multi-stage game development
Task execution plus role-specific review
Workflow review and completion
OpenGame ( Jiang et al., 2026b )
End-to-end Phaser game and asset creation
Build/test verification; companion benchmark executes the game
Engine validation; screenshots and runtime video studied as agent feedback
Development-agent benchmark
OpenGame-Bench ( Jiang et al., 2026b )
Finished web games
Headless browser execution and multimodal judging
Build Health, Visual Usability, and Intent Alignment scores
GameCraft-Bench ( Luo et al., 2026 )
Complete Godot games and replay traces
Replayed inputs, gameplay video and frames, hidden rubric
Core Mechanics, Content Depth, Functional Visuals, and Art scores
Table 1: Representative game systems and the role of evidence.
Agent
Responsibility
Owned artifact class
design
Author the multimodal GDD: the design specification plus linked concept mockups per screen and grounded crops of each element
Design and concept evidence
level-design
Materialize or revise authored levels; dispatched only when routing calls for them
Authored level set
scaffold
Project the multimodal GDD into a Phaser 4 shell plus the machine-readable contracts and deterministic tests that later gates read
Runtime shell, contracts, and tests
implement
Implement one scene’s gameplay and its feature tests; dispatched once per scene, serially
Scene implementations and feature tests
assets
Generate and register visual assets for one batch through the control plane, with per-asset metadata sidecars
Visual assets and metadata
asset-integration
Wire the finalized manifest into runtime preloads, animations, and display parameters, then seal the integration receipt
Runtime asset wiring and sealed receipt
Table 2: Specialized agents and their owned artifacts.
Figure 3: Three Deckbuilder scene contracts link multimodal GDD concepts to matching downstream runtime captures.
Figure 4: Animated-sprite production alternates bounded generation, checks, retries, and one post-join finalization.
Figure 5: One verification epoch. Checks run against the freshly sealed build and produce the evidence shown here, from a Minesweeper run. The gate and the critic then read the sealed report and either advance the run or return a blocker to its owner for a new epoch.
Category
Tools
Scope, with a representative tool
Lifecycle
9
Workspace leasing, preflight readiness, scaffold reproject receipts, and the transactional state journal — write_pipeline_state
Asset generation and validation
20
Image and video synthesis, matting, spritesheet packing, and the closed-loop producers that retry against VLM rubrics — produce_validated_sprite_action
Verification
16
The checkpointed verification DAG, its individual producers, and the gate runner — run_verify_dag
Concept grounding
3
Detect element boxes on an approved mockup and crop them into art seeds — detect_layout
Post-processing and matting
3
CPU compositing plus the two matting backends, remote and deterministic — rmbg2_matting
Video understanding and recording
3
Record a running game with sampled engine state, then analyze only that sealed recording — record_with_state
Table 3: MCP control-plane tools by category.
Figure 6: Single-writer mutation control accepts owned writes and routes rejected cross-component changes to the declared owner.
No.
Task
Family
No.
Task
Family
1
Minesweeper
Puzzle
11
Blackjack
Card
2
Snake
Arcade
12
Endless Runner
Runner
3
Tap Flyer
Arcade
13
Fruit Slicer
Action
4
Shmup
Action
14
Solitaire (Klondike)
Card
5
Match-3
Puzzle
15
Link Match
Puzzle
6
Tower Defense
Strategy
16
Gold Miner
Arcade
Table 4: Fixed 20-task paired corpus with game families, numbered in report-local sequence. Tasks 1–10 are listed on the left and 11–20 on the right. Task names describe gameplay genres and mechanics; they do not refer to any commercial game.
Metric
Output
Rubric question
Scoring
Requirement Satisfaction
[0,1]
What fraction of the observable prompt MUST requirements is satisfied?
Mean of all prompt-derived Core Mechanics, Content Depth, and Functional Visuals items.
Core Mechanics
[0,1]
Do controls, rules, state transitions, outcomes, and restart behavior work?
Weighted fraction of passed deterministic MUST-item checks.
Content Depth
[0,1]
Are requested modes, levels, progression, and terminal content present?
Weighted fraction of passed deterministic MUST-item checks.
Functional Visuals
[0,1]
Are required game state, HUD, action feedback, and outcomes visible?
Mean of direct evidence-grounded GPT-5.6 MUST-item scores.
Art & Presentation
[0,1]
Is the presentation coherent, readable, and visibly finished?
Mean of three shared GPT-5.6 visual-quality items.
Table 5: Test metrics and scoring rubric.
Mean ± standard deviation
Metric
GCA
OpenGame
Requirement Satisfaction
0.786 ± 0.151
0.656 ± 0.247
Core Mechanics
0.852 ± 0.144
0.655 ± 0.267
Content Depth
0.712 ± 0.286
0.651 ± 0.267
Functional Visuals
0.711 ± 0.192
0.667 ± 0.248
Art & Presentation
0.797 ± 0.099
0.827 ± 0.068
Table 6: Aggregate prompt-grounded results. Each system contributes one game per task and every game is scored three times, so a cell averages 60 evaluation runs over 20 games and its deviation spans tasks as well as scoring repeats. Bold marks the higher mean in each row. Core Mechanics and Content Depth derive from deterministic checks, Functional Visuals and Art & Presentation from grounded GPT-5.6 judgments, and Requirement Satisfaction averages the prompt-derived items across the first three dimensions. Null values are excluded, not converted to zero.
Reported quantity
GCA
OpenGame
Rubric ratings on a 1–5 scale (mean ± standard deviation)
Requirement Satisfaction
3.55 ± 1.36
3.28 ± 1.47
Core Mechanics
3.40 ± 1.31
3.16 ± 1.36
Content Depth
3.43 ± 1.36
3.28 ± 1.40
Functional Visuals
3.33 ± 1.35
3.29 ± 1.33
Art & Presentation
3.57 ± 1.13
3.03 ± 1.19
Table 7: Human study on 20 prompt pairs. Ten participants each rated all 20 pairs, giving 200 participant–game ratings per system. Bold marks the higher value in each row, and Overall is the mean of the five dimensions. Of the 200 preference choices 59 were “no preference”; excluding ties, GCA is preferred in 86 of 141 decisive choices (two-sided sign test, p=0.011 ).
Figure 7: Gold Miner formal-run evidence: retrieval, exact $500 credit, shop purchase, and level-2 transition in one continuous recording.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Tool
W
Purpose
ping
Health check: workspace and credential status.
init_workspace
Open one game directory under explicit create or resume intent; pins the inode and takes an exclusive repository-wide lease.
close_workspace
Detach the workspace and release its lease; gated tools then fail.
Table 12: Checks in iteration_ready , grouped by function.
Figure 8: Composite gate sizes, from the seven-check scaffold boundary to the 39-check release gate.
Key
Value
Effect
Iteration budget and evidence freshness ( pipeline )
min_iterations / max_iterations
1 / 2
Required fix rounds, and the hard cap
max_consecutive_blocks
5
Safety valve ahead of the budget
min_screenshots / min_unique_screenshots
3 / 3
Distinct screenshots required as evidence
staleness_tolerance_seconds
60
Clock-skew slack on freshness
max_analysis_age_minutes
30
Analyses expire after this
Appendix
Table 13: Default pipeline, gate, provider, and verification thresholds.
ID
Name
Condition for passing
I1
Identity match
Species, gender, era, art style, and notable features match the description exactly; no substitutions or embellishments.
I2
Hand-prop attachment
Every prop shows a visibly gripping hand or forearm; a prop merely near the body fails, however plausibly placed.
I3
Limb integrity
All expected limbs present, torso-connected, at realistic joint angles; none missing, extra, duplicated, or clipping.
I4
Foot anchor
Lowest visible point in the bottom 30% of the canvas; a floating or vertically centered figure fails.
I5
In-frame
Character and props unclipped, with at least 20% margin above the head and 15% each side for later motion.
I6
Background pure
Uniform solid chroma-key color; no gradient, shading, noise, secondary object, or border.
Appendix
Table 14: Image rubrics I1–I8, applied to sprite first frames.
ID
Name
Condition for passing
Framing group, judged on the raw video
R1
In-frame
Nothing is clipped in any frame, including peak motion: every swing, sweep, dash, and trail stays fully inside the canvas. Feet on the bottom edge pass.
R2
Scale locked
Size stays within roughly 10–15% of the first frame, judged late-against-first so creep is caught. Extending a limb passes; scaling the body does not.
Content group, judged on the matted output
R3
Identity preservation
Same colors, patterns, face, and proportions in every frame; no drift, swap, or morphing.
R4
Hand-prop attachment
Held items stay anchored to a hand or the body; nothing floats, detaches, or hovers without a visible grip.
Appendix
Table 15: Video rubrics R1–R7.
Dimension
Scale, from 0.1 through 0.3 and 0.5 to 0.8+
visual_quality
Incoherent; rough or accidental; intentional but uneven; profile-coherent, polished
gameplay_mechanics
No gameplay evidence at all; partially working; mostly working; all designed mechanics visible
completeness
Multiple major features missing; one or two missing; all present but rough; complete and refined
polish
No transitions or feedback; basic transitions; has juice; feels polished
asset_fitness
Missing, broken, or illegible; inconsistent or poorly integrated; runtime-fit but rough; cohesive and correctly integrated
Reproject deletion is receipt-mediated: scaffold seals its own stale list, and the orchestrator can only consume that receipt, never supply or widen the paths.
implement
Dispatch : one scene Required : scene_name , gdd_path , game_dir , genre , interaction_profile , quality_profile , level_model , url Optional : visual_evidence , on visual fix rounds only
scene , files_changed , tests , issues , design_change_requests
Appendix
Table 17: Phase-agent dispatch interfaces.
Field
Meaning
game_dir
Absolute game-project root; MCP tools take the same path under the same name
gdd_path
Absolute path to the game design document
genre
Startup-selected genre; downstream agents never re-route it
quality_profile
One of art_directed , responsive , pixel_art , minimal
level_model
One of authored , procedural , board , none
interaction_profile
Startup-selected physics and interaction route
Appendix
Table 18: Fields shared across dispatch interfaces.
Phase
Continuation rule
Design, level design, scaffold
No automatic retry. Design converts a bounded concept-generation failure into its documented warning fallback: complete layout and mechanics ship without the image.
Implement
No automatic scene retry; partial issues enter the fix loop, and resume targets only incomplete scenes.
Assets
Internal retries are bounded by configuration; a terminal partial may continue with only the missing keys.
Asset integration
One full dispatch first; scoped continuations are for confirmed timeouts only and run sequentially in the order config, preloads, animations, visual, verify.
Verify
Resume the same epoch from checkpoints after a confirmed interruption. Never restart completed nodes, and never open a fresh epoch for a transport failure. A new run starts only after the build changes or the previous one finalizes.
Critic
No automatic retry; a failed judgment blocks assessment rather than defaulting to either verdict.
Appendix
Table 19: Per-phase retry and continuation policy.
A game ships when a player can launch it, finish it, and return to a saved world, none of which a successful build guarantees, yet game-creation agents are mostly tested on one engine or on one-prompt browser games. We present CEGame-Bench, the first long-horizon benchmark for complete-game creation across five production engines: Defold, Godot, Roblox Studio, Unity, and Unreal Engine. Agents build each game from an empty project and a design document in runs averaging up to 295 steps and 65M tokens; the document doubles as a contract, and a hidden evaluator accepts a game only by real-input play. Under four scaffolds, agents earn most integration points but usually under half the gameplay points, and their games reach the ending in only 16% of runs, while references built by engineers with coding agents score near the maximum, which places this artifact-to-play gap in the agents. No two engines rank configurations alike (Kendall's tau 0.14-0.62), and the same game usually scores higher on Unity and Unreal Engine, which check more before play. We release Game Completion Boost, an agent skill for requirement-driven playtesting and repair that raises the execution-based score of three models by 14.1-24.3 points.
Zixuan Dong, Yanghai Wang, Hanyan Bian +8
National University of Defense Technology · Nanjing University · The University of Tokyo +2
Game generation is an emerging application of coding agents, requiring models to transform natural-language specifications into playable interactive systems. Unlike traditional coding tasks, game generation takes place within a game engine, where scripts, scenes, assets, rendering, and runtime interactions must jointly produce coherent gameplay. We formalize end-to-end game generation as the problem of producing a complete game artifact that realizes a specification through observable player-game interaction in a target environment. We argue that evaluating this setting requires three desiderata: Engine Grounding, Artifact Completeness, and Interactive Verification. We propose an interaction-grounded evaluation framework that assesses executable gameplay through replayed demonstrations and rubric-guided multimodal judging. We instantiate this framework as GameCraft-Bench, a benchmark comprising 140 Godot tasks across 15 game families. Evaluations of frontier coding agents show that end-to-end game generation remains highly challenging: the strongest agent achieves only 41.46%, and most agents score below 40%. Further analysis reveals that while agents often implement recognizable mechanics, they struggle to deliver complete games with sufficient content, functional visual feedback, and coherent presentation. See https://tongxuluo.github.io/gamecraft-bench-website for demos, code, and data.
Tongxu Luo, Rongsheng Wang, Jiaxi Bi +22
1The Chinese University of Hong Kong, Shenzhen · 2Shenzhen Loop Area Institute · 4USTB +5
Large language model agents have enabled software development through natural-language interaction, but extending this paradigm to end-to-end game development remains challenging. Building a playable game requires coordinating code, scenes, assets, animations, physics, and configurations, while allowing developers to validate their integrated behavior through interactive testing. We introduce VibeGame, a framework for creating complete 2D game projects from natural-language prompts and optional reference images. Its AI-native game engine represents project state as schema-validated structured text and provides frame-synchronous runtime control, enabling agents to inspect game states, execute actions, and verify behavior at their own pace. Built upon this interface, we design an Adversarial Agent Team of eight specialized agents to separate implementation from review, so that problems are found and fixed through multi-round loops. For each conducted and accepted project, we further introduce a training-free self-evolution mechanism to distill it into reusable skeletons, modules, and collaboration contracts for future development. We evaluate prompt-assets-to-game (PA2G), where agents are given an asset library, and prompt-to-game (P2G), where agents generate their own art. On a GameCraft-Bench subset, VibeGame in P2G consistently outperforms a range of agent harnesses in either P2G or PA2G with the same underlying model, including GPT-6.1-Sol.
Wenbo Hu, Ken Li, Jiazhe Wei +13
PRLab, Nanjing University · S-Lab, Nanyang Technological University