Which AI harness creates the most playable prototype: a test of P Agent, Claude Code, OpenCode, and DeepSeek

A game prototype from an AI harness is a software demo with controls, a combat system, enemies, and completion conditions that the agent assembles according to the user's instructions. The comparative test examined four options: P Agent, Claude Code, OpenCode, and DeepSeek's native harness, also referred to in the materials as Dipsik Hardness.

The test author named DeepSeek Harness the best according to the overall criteria. P Agent took second place thanks to its low cost and speed. OpenCode came third because of its long runtime and instability, while Claude Code took fourth place, even though it was the only game that could be completed.

The ranking does not demonstrate the universal superiority of the tools. It is the result of one author's scenario, one series of generations, and a subjective assessment of prototype quality.

How the AI harnesses were tested

The author compared more than just the appearance of the finished game. The test considered controls, combat convenience, the behavior of four enemy types, damage balance, animations, collision errors, launch stability, cost, token usage, and generation time.

First, each harness created an orc, then a simple warrior, a heavy warrior, an archer, and a shaman. The system then assembled a camp diorama, and the selected elements were used in the final game. This approach made it possible to test different levels of agent operation: an individual character model, a scene made up of several objects, and an interactive prototype with combat rules.

The assessments belong to the experiment's author. Phrases such as “best animations” or “poor balance” cannot be considered measurements on a single official scale, since the test did not define a formal set of automated metrics.

P Agent: convenient controls and overly harsh combat

P Agent created a game controlled with the WASD keys. The mouse handles sword attacks, viewing the scene, and interacting with objects, so the control scheme feels familiar for a three-dimensional action game.

The prototype's main problem is its balance. The hero has little health, the enemies are numerous, and ranged attacks quickly reduce the margin for error. There is no block or healing items, so a group of orcs leaves almost no room for tactics. In the author's assessment, defeating all the enemies under these conditions is practically impossible, although the game itself seemed fairly decent to him.

The shaman usually retreats and does not attack the hero. The author assumed that this behavior protects the shaman from being defeated early, although the test provides no confirmation of this interpretation. The orc with a spinning attack has a very large health pool, the archer runs away after taking damage, and the mage is effective at range.

The sword movements look quite good: blood sprays out and then disappears, while arrows, magical attacks, and other projectiles can hit the character at the same time. There is dynamism, but the lack of blocking and health regeneration turns active combat into a series of attempts to survive a volley.

Claude Code: the only completed prototype

Claude Code's game was the only one the author managed to complete, collecting the gold. This is the option's main practical advantage in the comparative test, although completing the game does not mean that it had no balance problems.

The character and animations looked somewhat weaker than those of the comparison option, but they retained an overall level of quality. Combat seemed slightly easier to the author because enemies attacked less often while moving. The mage threw magical projectiles, but could not hit the hero through the campfire.

The health pool was still insufficient for a relaxed fight against all eight enemies. Some enemies removed almost all of the character's health with a single attack, so Claude Code's advantage arose more from enemy behavior and the route through the level than from a well-tuned combat system.

In the overall assessment, the author noted an attractive and sufficiently detailed character, but gave negative marks for the animations and final balance. Claude Code took fourth place because it worked slowly, and the result seemed mediocre to the author more often than not.

OpenCode: interesting attacks and noticeable technical glitches

OpenCode used the left mouse button or the spacebar to control attacks. The sword strike looked more interesting than in some of the other options, and the author called the attack animations probably the best among the tested prototypes.

Stability was a weak point. The game occasionally froze, and the hero began the scene surrounded by a large number of enemies. The heavy orc dealt significant damage with a single hit, while the archer and mage almost never missed even a moving character, so the fight quickly became a test of survival.

The enemies also created visual problems. Blood remained suspended in the air after hits instead of disappearing. The character looked less detailed, and the author rated the enemy behavior and difficulty balance negatively, despite the successful individual weapon movements.

OpenCode took third place. It worked longer than any of the others, but produced slightly better results than Claude Code and looked more convincing in certain elements, especially its attack animations.

DeepSeek Harness: a stable result at high difficulty

DeepSeek Harness created a dynamic game with adequate attacks, but made the character more square-shaped. The mouse controls were vertically inverted, and the enemies sometimes interacted incorrectly with one another.

Combat centered on a high density of enemies and a lack of healing items. The heavily armored orc posed the greatest danger: it was difficult to knock down, and a single hit dealt substantial damage. The mage attacked effectively from a distance and did not engage in close combat, while the archer interfered with the fight and ran away after attacking.

The orc that spun around itself had a large health pool. Blood was present, sprayed after hits, and then disappeared. In the author's assessment, the game turned out to be highly dynamic, but problems with the animations and balance prevented him from calling it fully finished.

The author was unable to complete this prototype. He believes that healing items or less numerous enemy attacks would have made it noticeably easier to finish. In the final ranking, DeepSeek Harness took first place for its stable and generally good generations, although it cost more than the other options.

Tokens, cost, and game generation time

Below are the data recorded by the author during the experiment. These are not the services' current rates or a calculation of the cost of future projects. The final price depends on the model, mode, caching, number of iterations, and settings of the specific harness.

P Agent led in speed, while OpenCode took the longest. DeepSeek Harness cost approximately as much as Claude Code but completed the generation faster. Expenses cannot be compared solely by the number of cached tokens: caching affects the calculation differently, and the test materials do not provide a single method for recalculating all operations.

What the orc and diorama generation showed

The difference between the harnesses became apparent even before the game was created. P Agent made a recognizable orc with clothing, fangs, nails, and pronounced musculature, but the character had a strange face, misaligned pupils, tucked-in thumbs, and an unusual running animation.

Claude Code created a more menacing face with better-positioned fangs, but the face looked like a solid mass, and the character once again ran with its palms facing upward. OpenCode received a higher assessment from the author for its silhouette, facial features, and musculature, and its running looked the most natural. The flaw was clothing that stood away from the body.

The orc from DeepSeek's native harness had well-developed muscles, facial features, and body details, but its fangs were too large and the number of fingers was questionable. The running animation with the palms facing upward disappointed the author, although the character's appearance was expressive.

In the dioramas, DeepSeek Harness received the highest assessment for lighting and overall construction. The cart wheels were positioned correctly in the scene, and sparks, a flag, a forge, a table, and bowls of fire appeared. Shortcomings remained as well: the orcs collided with one another, and rotating the overexposed circle changed the shadows too dramatically.

The P Agent diorama had almost no noticeable defects, apart from the strange movement of the orc with the two-handed axe and occasional character collisions. OpenCode produced interesting light, shadows, campfires, and lights on the towers, but left problems with the roofs, palisade, cart, and collisions. The Claude Code option was the cheapest in this part, but contained an upside-down roof and an incomplete storage building.

Final ranking by task

• First place, DeepSeek Harness. The best combination of stability, speed relative to the result, and overall generation quality. Its weaknesses are related to the high cost, inverted controls, and excessive combat difficulty.
• Second place, P Agent. The fastest and least expensive option in the test. It creates a dynamic game with convenient controls, but requires serious manual adjustment of health, blocking, healing, and the number of enemies.
• Third place, OpenCode. Its strength lies in attack animations and individual visual solutions. Long generation times, freezes, nearly infallible ranged attacks, and visual bugs reduce the prototype's practical value.
• Fourth place, Claude Code. The only completed prototype, with the gentlest enemy behavior while moving. The tool lost out in duration, cost, and the stability of its visual results.

What will have to be refined manually

An AI harness can already assemble a playable foundation: a character, a scene, several enemy classes, ranged and melee weapons, simple hit effects, and a victory condition. Playability does not arise from one successful prompt, but from a combination of systems that the agent still often configures inconsistently.

The first stage of manual refinement should be balance. The developer needs to check the hero's health pool, each enemy's damage, the frequency of ranged attacks, the ability to block a hit, the availability of healing, and the safe intervals between attacks. If enemies attack simultaneously, even well-functioning animations cannot save the prototype from feeling unfair.

Next, collisions and navigation should be checked. Orcs should not get stuck inside one another, weapons must remain in their hands, blood should be removed after a set amount of time, and characters and scenery should not hang in the air. A final pass is also needed for the controls: mouse inversion, attack delays, and random freezes are immediately noticeable to the player.

Which harness should an indie developer choose? P Agent is suitable for an inexpensive, fast draft; OpenCode is interesting for experiments with attack animations; and DeepSeek Harness is the more sensible place to start when attempting to obtain a more stable overall result. Claude Code has a practical advantage in this test because its game could be completed, but one successful playthrough is not enough to consider the tool the best by every criterion.