filesystem-context

To manage large agent context that exceeds token limits, you can offload large outputs to the filesystem as scratch pads and return compact summaries to the context window. This approach provides efficient storage and retrieval strategies.

The best way to persist long-horizon task plans is to write them directly to the filesystem. This maintains task awareness over extended periods by storing plans as structured data files that the agent can retrieve selectively.

You can enable file-based communication for multi-agent coordination by using the filesystem as a shared messaging layer. Sub-agents write outputs and status updates to files, allowing other agents to dynamically discover and retrieve the context on demand.

Yes, you can dynamically load skills from files based on task requirements. Store skills as individual files in the filesystem and use search techniques like glob and grep to load the relevant skill instructions only when needed.

Dynamic context discovery requires combining standard filesystem search techniques such as ls, glob, grep, and read_file. These operations allow agents to navigate directories and selectively access full file contents only when required.

Yes, filesystem-based context management works without external dependencies because it relies entirely on standard filesystem operations and structured data storage. You only need knowledge of file handling to persist, retrieve, and offload context.