My last post ended with me wondering whether further harness improvements could really make the agents I was integrating into RISE more creative. It turned out that the agent building these scenes wasn’t failing for a lack of knowledge. When asked in isolation how to model a mermaid or a wizard it produced something convincing on the first try. Inside a live session though the same model produced a cone. A loaded context was consuming the attention that the work required and no amount of asking for creativity got that back.
So I stopped asking and instead over the last two weeks went into restructuring how the harness works. The harness first uses the model to ‘imagine’ the final scene and then feeds this back to the model during the creation and refinement process along with its context to drive the loop. I also added new expressive geometry and material capabilities to give the models more to work with. As a part of the harness guiding the planning process, it nudges the model to isolate its work on each piece of the scene and then place them, again a management of context. All of this made a remarkable difference:
| before | now | reference | |
|---|---|---|---|
| parts per figure | 1 to 2 | ~150 | 131 |
| objects in scene | 3 to 5 | 47 | 47 |
| physical light sources | 0 | all of them | 0 of 8 |
That last row is my favorite. The agent now lights the scene entirely with emissive surfaces, actual physically correct light sources which my reference image (made with an offline much more powerful model) doesn’t do.
Both scenes in the video below, an apothecary’s work table and that underwater prompt, were built end to end by the agent (I am using Gemini 3.7 Flash) from a single request. During my development Gemini 3.7 became available and I found it to be more capable than 3.6 Flash and also in the few scenes I have tried just slightly edging out GPT 5.6 Sol. My reference scene was created outside the harness in Claude Code with Fable.
There is still a ways to go. I’ve already implemented some additional materials (coming in my next posts) and I’ll also post a demo of how to iterate on the scene with the model. The ability for the model to generate complex objects like realistic animals or humanoid figures seems to be limited, even though when I examine the trajectories and look at the imagined image, they are very compelling. I think it still shows that there is a lot more to squeeze out from the harness and more verbs in the RISE language.
Aravind 