Dynamic context discovery is reshaping how coding agents operate, improving efficiency and response quality. Cursor's approach focuses on reducing static context and allowing agents to pull only necessary information as needed. This strategy minimizes token usage, which is critical given the limitations of context windows in these models. By avoiding the inclusion of excessive or potentially confusing information, Cursor enhances the agent's ability to generate relevant responses. Cursor implements this concept through several methods. For instance, long tool responses are converted into files, allowing agents to access output without overwhelming the context window. Instead of truncating data from shell commands or other tool outputs, Cursor writes them to files. This reduces the need for summarization and preserves important details. Another tactic involves referencing chat history during summarization, which helps agents recall specific details rather than losing them in a compressed format. The support for Agent Skills, an open standard, allows for dynamic context discovery of specialized capabilities. Skills are defined in files, enabling agents to access relevant tools efficiently. Cursor also reduces context bloat when accessing MCP resources by syncing tool descriptions to a folder, ensuring agents only load what's necessary. In A/B testing, this approach cut token usage by nearly 47% when using MCP tools, demonstrating significant efficiency gains. Lastly, Cursor treats integrated terminal sessions as files, eliminating the need to copy and paste outputs. This allows agents to query recent command outputs directly, making them more effective in troubleshooting. Overall, Cursor’s file-based methods offer a straightforward yet powerful way to enhance coding agent performance while managing context limitations effectively.