The AI thinks, the gate decides — how I made LLM code edits deterministic (and cut token usage 42 )

On September 10, 2026, I conducted a series of programming tasks using the same model, prompt, and repository. The tasks involved modifying a specific file in the repository's codebase. I ran the tasks three times using three different approaches: D-Engine, a deterministic pipeline, and two other AI coding tools. The results showed that D-Engine consumed significantly fewer tokens and completed the tasks much faster than the other tools.

The difference in token usage between D-Engine and the other tools was not due to the model, thinking mode, or the agent's toolbox, but rather the architecture of each approach. This discovery led to the development of D-Engine, a tool that separates the model's thinking from its ability to make changes to the code. The thought process of the model is confined to generating patches, while the local runtime is responsible for applying those patches.

This separation allows for deterministic and reliable code modifications while minimizing token usage.