filesystem-context

Write large tool outputs to scratch files and return a compact summary with the file path instead of the full content. The agent then uses grep or line-range reads to retrieve only the sections it needs, keeping roughly 100 tokens in context instead of thousands.

Save the plan as a structured JSON or YAML file on disk with objective, steps, and status fields. The agent re-reads the file at the start of each turn or after summarization to restore awareness of the objective and current progress.

Use it when tool outputs exceed roughly 2000 tokens, tasks span multiple turns, or multiple agents share state. Avoid it for single-turn tasks, when context fits comfortably in the window, or when file I/O latency is unacceptable.

Each sub-agent writes findings to its own isolated workspace directory, and the coordinator reads those files directly. This avoids the summarization degradation that occurs in multi-hop message chains and prevents write conflicts.

Self-modification can accumulate incorrect or contradictory instructions over time. Guard it with validation such as entry limits, value length caps, and periodic human review of persisted preferences.