Deciding how much autonomy to grant AI agents depends on the task, not just the model's sophistication. A framework based on task checkability and undoability helps determine appropriate delegation levels.

Tasks are categorized into four levels: Level 0 (Agent as assistant) for hard-to-check, hard-to-undo tasks; Level 1 (Human-in-the-loop) for hard-to-check, easy-to-undo tasks; Level 2 (Agent delegation) for easy-to-check, hard-to-undo tasks; and Level 3 (Self-driving mode) for easy-to-check, easy-to-undo tasks.

For tasks that are hard to check and costly to undo, like migrating sensitive code, break down the task and delegate less critical parts. For tasks hard to check but cheap to undo, use LLM-as-judge or define scoped goals for subjective evaluations. Easy-to-check, costly-to-undo tasks, common in development, benefit from automated checks and guardrails instead of human bottlenecks. Easy-to-check, easy-to-undo tasks are becoming more prevalent, with efforts focused on domain-specific models and robust context banks to enhance agent capabilities.