How to build agents with filesystems and bash

In exploring a novel approach to agent architecture, the text argues for simplifying the complex tooling traditionally used to feed information to agents by leveraging their inherent understanding of filesystems and bash commands. By replacing custom tools with filesystem and bash operations, agents can efficiently navigate and retrieve relevant data, as demonstrated by a significant cost reduction and improved output quality in a sales call summarization agent. The approach capitalizes on the extensive training large language models (LLMs) have in handling code and filesystem operations, suggesting that agents perform better when operating within a structured file-based context rather than relying on vector search or prompt stuffing, which can lead to imprecise results or exceed token limits. By structuring data in directories that reflect natural hierarchies—such as customer support tickets or CRM data—agents can precisely retrieve needed information through native operations like grep and cat, enhancing debuggability and security through sandbox isolation while minimizing maintenance. The text concludes by advocating for the adoption of this simplified architecture, emphasizing that the future of agent design might lie in the minimalism of filesystems and bash rather than elaborate custom solutions.