Development of Minimal Coding Agent Harness Begins Amid Increasing Complexity of Claude Code

Regarding Claude Code specifically, the author stated that while it was initially favored as a simple tool, recent versions have bloated to the point where 80% of the features are unnecessary. They expressed frustration with UI flickering and the fact that workflows break because system prompts and tools change with every release.

The author believes that "context engineering"—strictly controlling what is input into the model's context—is the most critical factor. They pointed out that existing harnesses make this control difficult because they inject information in the background that is not displayed in the UI.

Furthermore, the author argued for the need for a tool that makes all aspects of the interaction with the model inspectable and provides a clean session format that allows for post-processing. They also stated that existing APIs have developed bugs through organic evolution, which has degraded the developer experience.

To solve these problems, the author has begun developing their own coding agent harness based on the philosophy of "do not build it unless necessary." They intentionally chose a name that is difficult to search for to avoid increasing the user base, thereby preventing problem reports from flooding their GitHub issue tracker.

On the technical side, they built a unified LLM API based on four foundations: OpenAI's Completions API, Responses API, Anthropic's Messages API, and Google's Generative AI API. Since API behavior varies by provider, they implemented an abstraction layer to handle the specific characteristics of each company.