Many developers create complex tools to provide agents with information, but this often leads to fragile systems based on guesswork. A more effective solution involves using filesystems and bash commands. By switching to this simpler architecture, a sales call summarization agent reduced its cost per call from about $1.00 to $0.25 while improving output quality. The core idea is that large language models (LLMs) are already familiar with navigating filesystems, having been trained on extensive codebases. This familiarity allows agents to efficiently manage various types of data, such as customer support tickets and sales transcripts, by structuring them as files. In practice, the agent operates in a sandbox environment where data is organized into directories. When it needs context, it executes Unix commands to explore the filesystem, retrieves relevant information, and sends it to the LLM. This approach avoids the pitfalls of traditional context management techniques, such as prompt stuffing, which can exceed token limits, or vector search methods that may return imprecise results. Filesystems provide a structured way to manage data relationships, facilitating precise retrieval and keeping context minimal as the agent loads files only when necessary. The article provides examples of how different domains can map to filesystem structures. For instance, a customer support system can be organized with directories for profiles, tickets, and conversations, allowing the agent to quickly locate relevant resolutions. A sales call summary agent can analyze transcripts by referencing a clear file structure that includes call metadata and prior conversations. This method allows the agent to operate efficiently, searching for specific patterns without relying on predefined retrieval logic. Using bash commands like grep and cat leverages the native capabilities of LLMs, ensuring that the agent can utilize familiar tools without extensive retraining. The architecture is adaptable; as models improve, so do the agents. This setup also enhances security by isolating the agent's operations from production systems. Overall, this approach simplifies agent development, as creating a filesystem structure can often replace the need for complex retrieval pipelines.