11 hours ago
- Agent autonomy depends on two factors: ease of checking the agent's work and cost of undoing mistakes.
- Four levels of autonomy: Level 0 (assistant) for hard-to-check and costly tasks; Level 1 (human-in-the-loop) for hard-to-check but cheap to undo; Level 2 (agent delegation) for easy-to-check but costly to undo; Level 3 (self-driving) for easy-to-check and cheap to undo.
- To increase autonomy, break complex tasks, use LLM-as-judge for subjective checks, enforce guardrails with code, train domain-specific models, and build expert-level context banks.
- Scale is not a primary factor; focusing on task-level autonomy naturally handles scale.