Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

01 / AI · Analysis

SceneAgent separates scene-building knowledge from engines

SceneAgent separates reusable scene-building guidance from engine integration, making execution cost part of the case for reuse.

AI-generated editorial illustration for this article

SceneAgent researchers reported evaluating their game scene generation framework on 69 prompts under shared asset and evaluation conditions.

The paper described a migration problem for game developers: planning and construction tools can remain tied to the original engine. Exporting the finished scene alone does not resolve those dependencies.

The researchers presented a design that separates scene state and reusable construction knowledge from engine integration. They reported benefits from reusing guidance alongside the additional runtime it requires.

Scene-building knowledge outside the engine

According to the paper, SceneContext is a shared structure that maintains the state of a scene during construction. New construction agents query that state rather than inherit the full conversation from earlier work.

The paper describes Layout skills as placement instructions stripped of asset identifiers. The framework extracts these instructions and guidance about content meaning from human-approved work records. It retrieves and recombines that guidance for subsequent generation tasks.

The researchers reported retaining the shared construction procedure and guidance library when deploying the framework in another engine. They replaced the engine integration components. The design isolates coordinate conventions, asset connections and export in a separate integration layer.

The execution cost of reusable guidance

Together AI reported that evaluation performance fell after it merged its agent instructions, then recovered after revisions. Its company-published pass rates were 75% for the initial combined version and 97.2% for the revised version.

Together AI identified excessive context as a problem with the initial version. It shortened the instructions loaded for every task and changed the guidance so agents open only the reference files they need. It also corrected outdated product information and errors in example code.

Together AI stated that its evaluations at different points used different models and tasks, so they were not a controlled comparison. Those differences limit how much of the change in pass rates can be attributed to revisions to the instructions alone.

SceneAgent's researchers reported that reusable guidance and the multi-agent procedure improved alignment with requested scenes and scene completeness, while adding runtime. For game developers assessing reuse across engines, the design therefore calls for weighing retained construction knowledge against execution cost.