filesystem-context

You can manage agent context by offloading large tool outputs and persisting plans to the filesystem. This approach leverages file storage to prevent context bloat and enables dynamic information retrieval using tools like `ls`, `glob`, `grep`, and `read_file`.

The best way to store large tool outputs is writing them to files and returning a summary with a reference. Agents can then selectively retrieve specific data later using grep or file reading tools, keeping the active context lean and relevant.

Yes, agents can use the filesystem for sub-agent communication by writing findings to shared locations. This file-based exchange facilitates inter-agent coordination without relying on direct context window sharing.

You can load skill definitions dynamically from files on demand. This dynamic loading approach ensures that the static context remains lean by only retrieving and applying skill definitions when they are explicitly required.

Persistent planning works by saving agent plans to files, ensuring they remain accessible across long-running tasks. This prevents plan loss even after context window resets, allowing agents to resume operations seamlessly.

Dynamic context discovery utilizes tools like `ls`, `glob`, `grep`, and `read_file` for filesystem context management. These tools enable agents to navigate directories, search file contents, and retrieve specific data references efficiently.