filesystem-context

Agent context management prevents context bloat by offloading large tool outputs to files on the filesystem. The agent returns only summaries and references to the active context window, keeping it clean and focused on the current task.

Filesystem persistence stores agent plans and state in files for long-horizon tasks. Agents dynamically load relevant information on-demand using tools like `ls`, `glob`, and `grep` to retrieve specific insights later without exceeding context limits.

Sub-agent communication is facilitated through shared file workspaces on the filesystem. Agents write intermediate analysis results and share information via files, enabling dynamic context discovery and retrieval without passing large payloads directly.

Dynamic context loading retrieves skill instructions only when needed. Agents use filesystem tools to discover and read relevant instructions on-demand, preserving active context space and preventing unnecessary bloat from loading all instructions initially.

To save tool output, use the filesystem-context skill to write the output of the 'web_search' tool to a file named 'search_results.txt'. You can then use `grep` or `read_file` to retrieve specific insights from that file later.

Using the filesystem as a scratch pad is not ideal for operations requiring immediate real-time memory access without file I/O overhead. It is designed for overcoming fixed context window limitations through dynamic loading rather than serving as a high-speed in-memory cache.