agent-supervisionlisted

# Supervising running agents A subagent that misreads its brief burns its entire budget before anyone finds out. Judging output afterwards is necessary and insufficient — by then the cost is paid and the wave is late. Three mechanisms, and **they are not equally strong**. Say which is which whenever you report on supervision; claiming cooperative control is structural is the failure mode this skill exists to prevent. ## 1. The intent contract — structural Every agent emits this **before** any substantive work: ``` ## INTENT-CONTRACT files-i-will-touch: <explicit paths, or NONE> index-rows-i-will-cite: <entity ids / precedent ranks> questions-i-am-answering: <1-3 lines> out-of-scope-for-me: <what I will NOT do> checkpoints: <where I will re-read my inbox> ``` It costs almost nothing and it converts drift detection from a judgement call into a comparison. **An agent that touches a file absent from its own contract has drifted by its own declaration** — no interpretation required. Write each contract to `<run-dir>/supervision/intent-<agent>.md` so it survives the agent's context. ## 2. The intervention inbox — cooperative Every agent's brief includes: *"At each checkpoint you declared, re-read `<run-dir>/supervision/INBOX-<your-name>.md`. If it exists and is non-empty, it overrides your current direction. Obey it, or STOP and report why you cannot."* Three directives: - **`REDIRECT`** — the goal is right, the path is wrong. Correction plus reason; the agent adjusts and