How agents can use filesystems for context engineering

AI agents often fail due to poor context engineering, which is the process of supplying the right information to the model's context window. Common challenges include retrieving too many tokens, needing more context than the window allows, and difficulty finding niche information. Providing agents with access to a filesystem offers a single interface to store, retrieve, and update a virtually infinite amount of context. This allows an agent to use the filesystem as a scratchpad for large tool outputs, a place to store plans, and a way to search for information using tools like grep, complementing traditional semantic search. This method also enables agents to learn over time by writing new instructions or facts to files based on user interactions.