Files Are All You Need

Agents are increasingly using files as a primary interface for managing and interacting with unstructured context, a trend largely propelled by coding tools like Claude Code and Cursor. These files serve multiple purposes: storing and retrieving long-running conversation histories, accessing large amounts of external context, and providing skills to agents, thereby bypassing the need for multiple MCP tools. The file interface allows agents to manage various file types, including plain-text, documents, and images, using CLI commands and tools like Read for effective context management. The shift towards file-based systems enhances agents' capabilities by allowing them to dynamically search and retrieve context, which can outperform traditional semantic search methods, particularly on smaller document collections. The introduction of "Skills" as file-based resources further equips agents by enabling them to learn and perform tasks through code execution. Despite these advancements, challenges remain in processing non-plaintext documents and scaling file search operations. The file-based approach is emerging as a promising mode for agent interaction, aligning with the trend of programming through natural language and paving the way for future developments in generalized computing use.