Koda presented the results of testing coding agents on the DeepSeek V4 Flash model

Developers of Koda conducted a comparative performance testing of popular automation programming tools. During the experiment, the DeepSeek V4 Flash model was launched in four different execution environments to evaluate the effectiveness of solving applied problems.

The measurement results showed that the proprietary Koda harness demonstrated the accuracy of task execution at the level of 52.2%. This indicator allowed the tool to outperform alternative solutions, including Kilocode, Claude Code and OpenCode, which showed lower results in similar testing conditions.

As part of the study, the authors analyzed how the architectural features of the execution environment affect the final performance of the LLM. The main attention was paid to how the agent interacts with the file system and project context, which directly correlates with the success of completing tasks for writing and debugging code.

The difference in the final scores between the tested tools is due to the methods of processing requests and managing the sequence of actions of the agent. The use of specialized harnesses allows the model to more effectively use its computational resources, minimizing the number of redundant operations when working with the code base.

This study highlights the importance of choosing an infrastructure environment for deploying coding agents. Optimizing the interaction between the model and the execution environment becomes a key factor in improving the quality of automated development in real projects.