Context management is critical in agentic coding, where adeptly framing and managing context can unlock significant value. Users must provide coherent direction, but the orchestration layer—called a harness—plays a vital role in translating user intent into useful context for large language models (LLMs). The article emphasizes that even the smartest models falter with poor context, while clear, high-quality context can elevate their performance. The upcoming developments in harnesses, which work in tandem with model capabilities, promise to optimize context management and enhance output quality. The piece highlights the rise of "Plan Mode," a feature that allows LLMs to break down tasks and create structured plans. Since the introduction of Plan Mode with Claude Code, there's been a noticeable uptick in performance across various coding agents. Current features in Plan Mode are basic, often producing mediocre plans. However, there's an expectation for future iterations to improve significantly, incorporating user interaction and advanced planning techniques. The author advocates for a more sophisticated approach to planning, drawing comparisons to existing tools like Repoprompt’s Context Builder, which offers a deeper level of rigor in planning. Another key point is the blend of semantic search and traditional grepping as a method for enhancing search performance within codebases. The article discusses the Context7 Model Context Protocol (MCP), which indexes coding documentation for quicker reference, improving coding accuracy. However, it argues that simply indexing isn't enough. There needs to be a system for real-time retrieval of relevant documentation as needed. This is especially important for models that may not have the latest information. The author believes that as teams focus on improving documentation retrieval, LLMs will become much more adept at finding and utilizing the right information to minimize errors.