The author reflects on three years of using large language models (LLMs) for coding, starting from basic tools like ChatGPT to more advanced coding agents such as Claude Code. Initially, Claude Code met the author’s needs, but recent updates have introduced unwanted complexity and broken workflows. The author values simplicity, preferring tools that don’t overwhelm with features he doesn’t use. Frustrated by existing harnesses that inject unnecessary context and lack transparency, the author decided to build a new coding agent harness, aiming for a minimalist and opinionated design. The project consists of several components: pi-ai for a unified API across various LLM providers, pi-agent-core for managing tool execution, pi-tui for a terminal UI framework, and pi-coding-agent to connect everything. Key challenges included addressing the differences between APIs from providers like OpenAI, Anthropic, and Google, particularly around features such as token tracking and reasoning outputs. The author emphasizes the importance of context management when switching between providers and devised a system to handle context transfer. The tools also support browser functionality, which is beneficial for creating web interfaces. In the process, the author highlights the inconsistencies among providers, especially in terms of how they report token usage and manage session states. For example, while some providers offer straightforward token counts, others complicate accurate tracking. The author’s solutions, while not exhaustive, have proven effective for personal projects. The focus remains on building only what’s necessary, aligning with the author’s preference for straightforward, functional tools.