filesystem-context

To reduce agent prompt bloat, offload large tool outputs to the filesystem by writing them to a scratch pad file and referencing the file path in the agent context for on-demand retrieval. This keeps the core prompt lean while preserving data.

The best way to persist multi-step plans across context refreshes is using plan persistence features that save long-horizon task states to the filesystem. This ensures continuous task tracking without losing progress during context resets.

Yes, you can use filesystem-backed memory and sub-agent workspaces for cross-agent information sharing. By storing shared state in files, multiple agents can dynamically discover and retrieve context using grep and read_file style operations.

You need dynamic context discovery for long-running agent tasks when oversized tool outputs and long-term state exceed the prompt window. Offloading this data to files allows the agent to load only relevant content dynamically when needed.

Yes, this approach implements safe self-modification and guardrails for evolving agent preferences. It allows agents to securely update their behavioral state stored in the filesystem while maintaining operational safety boundaries.