Per-action autonomy means the level of automation is chosen task by task rather than all at once. For each kind of action you pick one of three settings: let the agent do it automatically, have it draft the action for your review, or leave it off entirely. Sending a routine tenant reply, retrying a failed rent payment, and dispatching a vendor can each sit at a different level.
This granularity is what makes autonomy safe to adopt. Instead of a single 'let the AI run everything' switch, you start by automating only the low-stakes, easily reversible tasks, keep sensitive ones in draft-for-review, and turn the rest off. Each setting is recorded, so it is always clear what the agent was and was not allowed to do.
Per-action autonomy is the practical on-ramp for trust: the design intent is that you widen the settings you are comfortable with as the record proves itself, not that you hand over blanket authority on day one. It sits alongside bounded autonomy, which adds spend and approval thresholds within each automated action.