What is an Agent Autonomy Levels?

Borrowed loosely from the self-driving-car classification of automation levels, agent autonomy is rarely all-or-nothing in practice. At the low end, an agent only drafts suggestions for a human to send; a step up, it acts automatically on routine, low-risk tasks but escalates anything unusual; higher still, it operates independently within a defined scope and only reports afterward; and at the top, it can even adjust its own strategy with minimal oversight. Most production deployments deliberately sit in the middle tiers, because full autonomy without escalation paths is where agents cause the most damage when they encounter an edge case they misjudge. The autonomy level should be configurable per task type, not a single global setting — an agent might operate with high autonomy drafting internal reports and low autonomy when the action touches money or external customers. Raising an agent's autonomy level is a decision that should be deliberate and reversible, not a default that creeps upward over time.