How AI harnesses build game dioramas: a comparison of P Agent, Claude Code, OpenCode, and DeepSeek

A comparison of AI harnesses for game dioramas shows how differently the same DeepSeek Q4.1 Flash handles a large scene. In this test, the AI harness passes the project files to the model, starts generation, checks the result, and continues working on the scene according to a given plan. The author compared P Agent, Claude Code, OpenCode, and DeepSeek’s native harness, which is also called Dipsik Hardness in the test materials.

The winner depends on the task. P Agent assembled the most cohesive camp at a moderate cost, OpenCode stood out for its lighting and shadows, while Dipsik Hardness delivered the most convincing overall result. It turned out to be the most expensive and did not eliminate all errors. Claude Code cost less than the other options but fell behind in visual coherence.

How the game diorama test was conducted

First, the author created an orc in each harness, then added a simple warrior, a heavy warrior, an archer, and a shaman. After that, each tool had to assemble one game diorama from the previously created characters and environment. At the final stage, the neural network combined the best elements from different generations into a single scene. The author did not assemble it manually.

The evaluation included four parameters: scene quality, cost, duration, and the number of noticeable defects. The author separately checked lighting, object assembly, animations, weapon placement, character collisions, and the plausibility of the environment. The results relate to a single test scenario. They do not prove that one harness is always better than another in every project.

The expected diorama composition was fairly complex. The camp was supposed to include campfires, houses or storage buildings, a forge, watchtowers, fences, a cart, animals, a flag, and orcs that could move interactively. Such a scene tests not only the model’s ability to write code. It also needs to maintain spatial relationships between a large number of objects.

P Agent: the most cohesive camp and collision problems

The P Agent diorama cost 76,6 cents, and its creation took 48 minutes 38 seconds. The process used nearly 209 thousand input tokens and 339 thousand output tokens. The amount of cached data was nearly 1 million tokens. The scene included all previously created orcs, a camp with a campfire, houses or storage buildings, a forge, towers, and grazing animals.

In the author’s assessment, P Agent assembled the most cohesive composition without a large number of immediately noticeable visual errors. The main elements were in their proper places, and the camp felt like a unified game scene rather than a random collection of objects. Its weak point was the animations and physical interaction between characters: the orcs sometimes collided, and the orc with a two-handed axe moved strangely.

The result was strong for a rapid visual prototype. The developer gets a scene that can already be shown to the team or used as a basis for further manual refinement. This is not enough for a released game: collisions, weapon grip, and NPC behavior require separate checking in the engine.

Claude Code: low cost versus assembly errors

Claude Code completed the diorama in 52 minutes 28 seconds. The cost was 45 cents. Input token usage reached nearly 372 thousand, output token usage was nearly 374 thousand, and the cache was about 17,45 million tokens. In terms of cost, this option was the cheapest of the four dioramas.

The low cost did not come with better quality. The author immediately noticed an upside-down roof, and one of the storage buildings looked incomplete. The animals looked more like dogs than boars, making the camp’s visual concept less convincing. The remaining elements generally resembled P Agent’s scene, but individual errors were more conspicuous.

Claude Code is suitable for an inexpensive rough pass when a developer needs to quickly test an object-placement idea or obtain a foundation for further work. Based on this test, the author did not choose it as the final diorama generator. The two stars reflect the low cost but also the noticeable quality gap.

OpenCode: expressive lighting and the longest generation

OpenCode created the diorama in 83 minutes 56 seconds. Usage amounted to 631 thousand input tokens and 214,5 thousand output tokens, the cache reached 103 million tokens, and the cost was 53 cents. Generation took longer than with P Agent and Claude Code, although the price remained relatively low.

The author named lighting as OpenCode’s main advantage. Shadows are noticeable in the scene, the campfires are burning, and lights appeared on the watchtowers. The orcs move around the camp, making the diorama feel more alive, especially in night mode. The lighting here does more than decorate the scene: it highlights the forge, campfires, and individual areas of the camp.

Errors remained. One storage building looked empty or difficult to distinguish because of the lighting, while another was filled. The orc with a halberd again held the weapon incorrectly. The characters collided, stopped, and stumbled over one another. In one scene, the cart was positioned incorrectly, and in the author’s opinion, the wheels should have been placed differently.

The night scene concealed some shortcomings but revealed others. Some fences appeared to float above the ground, and part of the palisade extended beyond the diorama’s boundaries. Such problems are sometimes unnoticeable in a beautiful screenshot. In an interactive game, they quickly become obvious when rotating the camera or moving the character.

Dipsik Hardness: the best overall assembly, but with caveats

The Dipsik Hardness diorama used 248 thousand input tokens, 441 thousand output tokens, and 88,8 million cached tokens. The cost was about 96 cents, and generation time reached 74 minutes 53 seconds. By these measures, the option was the most expensive and one of the slowest.

The author rated this scene as the most adequate of those considered. The lighting and overall assembly looked somewhat better, the wheels were installed correctly, the fire had noticeable sparks, and the camp included a flag, a forge, a table, and elements resembling chairs. The storage building, forge, and bowls of fire also looked good enough.

Dipsik Hardness had one notable defect of its own: an overexposed circle that caused the shadows at object intersections to change too dramatically when the scene was rotated. The orcs continued to collide with one another, and their movement across the map raised questions. At certain moments, one orc did hold the halberd correctly. This demonstrates an incomplete but noticeable ability of the harness to correct spatial errors.

Dipsik Hardness assembled the final diorama from elements of different generations. The author did not join the objects manually. This approach is important for evaluating autonomous operation, but it also increases the risk of inconsistencies: in the final scene, the table had been disassembled, while some earlier defects remained.

Which errors recurred in the procedural scene

Recurring defects show that generating a game diorama consists of several different tasks. The model may correctly understand the overall composition but make mistakes in geometry, animation, or collision rules.

• Character collisions. Orcs sometimes collided, stopped, and stumbled over one another. This is tolerable in a static demonstration, but a game requires navigation, colliders, and obstacle-avoidance logic to be configured.
• Weapon grip. The halberd, two-handed axe, and other items did not always align with the characters’ hand positions. The model created visually similar weapons but did not ensure a correct connection between the object and the character’s bones.
• Roofs and storage objects. Scenes contained upside-down and strangely assembled roofs, while one storage building looked unfinished or empty.
• Cart and wheels. In one diorama, the cart was positioned incorrectly. Dipsik Hardness corrected the wheel placement, but the other harnesses did not always manage this kind of spatial relationship.
• Fences and palisade. Night scenes concealed some visual errors, but some fences floated in the air, and a section of the palisade extended beyond the diorama’s boundaries.
• Lighting. OpenCode created expressive shadows and lights, but the lighting made one storage building difficult to distinguish. Dipsik Hardness produced an overexposed circle where the shadows changed sharply as the camera rotated.

What happens if only the beautiful frame is evaluated? The ranking will shift in favor of OpenCode and Dipsik Hardness: their lighting and shadows make the scene more expressive. Checking movement, collisions, and weapon correctness narrows the gap between the harnesses. All four options exhibit problems.

Comparative diorama table

Which harness to choose for a visual prototype

For a rapid camp prototype, the author’s test places P Agent first for its combination of speed, price, and result coherence. It did not provide the best lighting, but it assembled the scene without a large number of noticeable defects and completed it in less than 50 minutes.

Choose OpenCode if atmosphere, night lighting, and expressive shadows are more important for the prototype. The developer will need to allow more time to fix collisions and weapon placement, as well as objects that are difficult to make out in the dark.

Claude Code is a rational choice for an inexpensive proof of concept. Its result should not be treated as a finished diorama: upside-down roofs, an incomplete storage building, and unsuitable animals require manual checking from the very first pass.

Dipsik Hardness is suitable for a scenario where overall assembly, lighting, camp details, and an attempt to autonomously combine the best elements from different generations are important. The author awarded it four stars but did not call the result unequivocally superior to the others, since the harness was the most expensive and behavioral and geometric errors remained.

The final choice depends on the goal. P Agent is more practical for the first pass, OpenCode is more interesting for an atmospheric scene, Claude Code is cheaper for a rough check, and Dipsik Hardness makes the strongest overall impression in this particular experiment. None of the options eliminates the need to manually check collisions, navigation, weapon grip, roofs, fences, and NPC behavior before using the diorama in a game.

FAQ

Which AI harness turned out to be the best for a diorama?

According to the experiment’s author, the Dipsik Hardness diorama was the best, but P Agent looked more practical in terms of price, time, and scene coherence. This is the conclusion of a single comparison, not a universal ranking.

Which harness was the cheapest?

Claude Code produced the cheapest diorama. Its cost was 45 cents, but the author also noted an upside-down roof, an incomplete storage building, and animals resembling dogs.

Which one handled lighting best?

OpenCode stood out with burning campfires, tower lights, and noticeable shadows. Dipsik Hardness also received a high rating for lighting and overall assembly, but it developed an overexposed circle with sharp changes in the shadows as the scene rotated.

Can such a diorama be used in a game immediately?

No, the result should be considered a visual prototype. Before using it in a game, collisions, character navigation, weapon placement in the characters’ hands, the position of wheels, roofs, and fences, as well as NPC behavior when they collide with one another, must be checked.