Co-founders of the gaming accelerator Narwhal, Dmitry Filatov and Grigory Bortnik, attempted to determine how effectively artificial intelligence assists in game development. To do this, they conducted an experiment—initially having AI independently develop a web game about an indie studio producer The Last Game Producer, and then having AI play it.
The Last Game Producer
- According to Filatov and Bortnik, the first attempt to create The Last Game Producer ended in failure. They hoped one very detailed prompt would suffice for the game, but the hypothesis did not hold. Although the developers tried to thoroughly describe in the original prompt the references, mechanics, and many other aspects of the desired title, “hundreds of small game design junctions” still remained, in which the AI began to independently come up with solutions—often unsuccessful and plainly dull. Moreover, implementing changes in the project with this approach proved too difficult. Although the generated build was functional, the co-founders of Narwhal abandoned it and restarted development.
- During the second attempt, Narwhal resorted to the classic game development approach with prototyping and iterations.
- According to the developers, the final result was significantly influenced by their level of expertise in certain areas. For instance, the AI performed well with programming and localization of The Last Game Producer, as Filatov and Bortnik knew exactly what to expect from the neural network. However, they faced difficulties with sound design, an area where the experimenters lacked understanding.
- Problems also arose when aligning the text and visuals: at one point, the AI forgot to draw a cactus on the main character's desk in The Last Game Producer, even though it was mentioned in the scenario. Furthermore, the co-founders of Narwhal failed to conceal the use of AI in the graphics. Despite using different models and complex post-processing, all testers were able to recognize the generated images.
- At some point, Filatov and Bortnik had to create a shared task tracker in GitHub Issues for the AI agents used in the development of The Last Game Producer. They also delegated routine code proofreading to autotests and script linters, but retained the semantic check themselves. This allowed them to expedite the processes.
- In total, the creation of The Last Game Producer took about four months of evening work. According to Filatov and Bortnik, without AI, a single developer would have needed a year for the same task.
- During the experiment, 128 releases and 787 commits were made. The final project retained 28,000 lines of JS/CSS and data, although the AI wrote more than 1.22 million lines overall. The scale of the development was marked at 1.5-2 billion tokens, and the cost was estimated to range from $2,300-$3,000 to $7,000-$9,000, depending on the Opus 5.5 model tariffs.
- Once The Last Game Producer was ready, the developers ran three neural networks in the game, allowing them to step into the role of the main character—the indie producer. Each AI ultimately adopted different strategies. Notably, Claude Sonnet 5.5 focused on maximizing profits and pressured developers for potential hits, achieving a better result than average human players. Meanwhile, GPT-5.6 Terra proceeded cautiously, opting for safe contracts, while Gemini 3.8 Flash protected the team and product.
As a result of the experiment, Narwhal co-founders concluded that AI still cannot completely replace specialists, although it accelerates game development. While neural networks handle individual tasks fairly well, the quality of the outcome still relies on human expertise.