SceneAgent framework aims to make game scene generation portable across engines
A new arXiv paper describes a multi-agent system that separates scene state, reasoning and reusable skills so a generator can move between game engines.
In plain words
The authors present SceneAgent, a system of cooperating software agents that builds game scenes. Its core idea is to keep the scene's state, the agents' reasoning and the learned building skills in separate layers, so switching to a different game engine does not require rebuilding everything. The paper also introduces GameSceneBench, a set of 69 scene-building prompts used for testing.
What was announced
SceneAgent uses a canonical SceneContext that maintains the evolving world state, while backend adapters handle engine-specific details such as coordinate conventions, asset bindings and export. Structured handoffs allow new Builder sessions to query this persistent state without carrying over the full construction transcript, according to the abstract. Across generations, human-approved traces are distilled into reusable Layout and Content skills that can be retrieved and recombined.
The authors report that on GameSceneBench, under a common asset and evaluation protocol, SceneAgent achieved human tie-adjusted preference scores of 78.2-89.4% against five baselines. Ablations attributed gains in prompt alignment and scene completeness primarily to reusable skills and the multi-agent protocol, at the cost of additional runtime. A deployment in a second engine reportedly preserved the shared construction protocol and skill library while replacing backend integration components. These figures come from the authors' own evaluation and have not been independently verified.
What it does not show
- The evidence is limited to the paper's abstract; full methodology, baselines and engine details are not covered in the supplied passages.
- Reported preference scores and ablation findings are the authors' own results and do not constitute independent confirmation.
- The abstract does not name the second engine used in the cross-engine deployment or specify availability of the framework.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.