Graduated autonomy
Governanceautonomy levels · staged autonomy
An explicit ladder of agent permissions: suggest, execute with approval, execute and notify, execute silently. Each rung is earned with evidence — green evals, clean traces — and can be walked back the same day if quality slips.
Graduated autonomy avoids the two bad calls: full autonomy on day one, or suggestion mode forever. It is defined per action type, not per agent: the same agent can execute refunds below a threshold and only suggest above it. Each level carries promotion criteria (eval accuracy threshold, minimum case volume, reversal rate) and automatic demotion rules. This definition is Implementa's own editorial formulation. It gathers practices that circulate in the market under labels like human-in-the-loop or autonomy levels, but the four-rung ladder with evidence-based promotion and automatic demotion is our way of framing it, not an industry consensus.
How it differs from
- Human-in-the-loop
- Says a human is involved; says nothing about how or when they hand over the wheel.
- Feature flag
- Turns a capability on or off; graduated autonomy modulates permission to act.
Examples
- A collections agent that sends reminders on its own but needs approval to write off an amount.
- A content agent that publishes to draft for 30 days, then goes live if the eval stays green.
FAQ
- Who approves a level-up?
- The process owner on the client side, with the eval numbers in front of them — not the technical team on its own.
- Can you go back down?
- It must be possible live, and automatic once the failure threshold is crossed.