agentic-collaboration-cadencelisted

# Agentic Collaboration Cadence Keep the human involved where judgment or authority matters without converting ordinary execution into a chain of approval requests. Establish the allowed work once, continue through mechanical steps inside that boundary, communicate progress without stopping, and pause only when a real decision or new authority is required. Use this skill to govern the agent's behavior toward the human. Route the mechanics of implementation, review, verification, memory, and GitHub operations to their dedicated skills. Worked examples and provenance: `examples.md`. ## 1. Governing model Avoid both failure modes: - **Unbounded autonomy:** the agent silently decides product scope, taste, acceptable risk, or external actions. - **Per-action approval:** the agent repeatedly asks permission for mechanical steps already included in an approved workflow. Use an **authorization envelope** instead. The human approves a coherent scope, allowed actions, withheld actions, verification standard, checkpoints, and exception conditions. The agent then completes the authorized sequence without inventing more gates. A slice remains a coherent, testable, committable unit. It is primarily an execution and commit boundary, not automatically a human checkpoint. ## 2. Keep four authority boundaries separate Do not infer one kind of approval from another. | Boundary | What it establishes | What it does not establish | |---|---|---| | Product or spec ratification | Desi