Assign Autonomy Levels: Speed vs Responsible AI
Autonomy is a dial, not a switch. This topic walks through the standard Suggest → Review → Auto-apply → Autonomous ladder, explains which guardrails are non-negotiable at each rung, and shows how to trade speed for Responsible AI controls without collapsing into either extreme.
Assign Autonomy Levels: Speed vs Responsible AI
Autonomy is a dial, not a switch. The exam will frame the trade-off as speed of delivery against Responsible AI controls — and the right answer is rarely either extreme. The right answer is the lowest autonomy level that delivers the leverage you need, with the controls that level demands.
The standard ladder
| Level | Name | What the agent does | Required guardrails | | --- | --- | --- | --- | | 1 | Suggest | Proposes; never executes | Audit log | | 2 | Review | Executes against a draft artifact (e.g., PR branch) | Audit log + reviewable artifact | | 3 | Auto-apply (sandboxed) | Executes inside an isolated boundary | Audit log + reviewable artifact + automated eval suite | | 4 | Autonomous | Executes and ships to production | All of the above + runtime monitoring + rollback path |
Notice the controls accumulate — you do not trade safety for speed, you trade per-call human attention for stronger automated signals.
Try the dial
Autonomy dial
Pick the right level for each kind of action, and check the guardrails that must be in place.
Exam tip: when a question asks you to raise autonomy, the correct answer almost always adds an automated signal (eval, monitor, rollback) — not removes a human.
Where this shows up on the exam
Expect a scenario with a slow human-approval bottleneck and a candidate set of changes. The wrong answers will either (a) remove the human entirely with no new controls or (b) keep everything in Suggest mode. The right answer adds the next signal up the ladder.
Key terms
- Autonomy level
- A discrete rung on a ladder that pairs how much the agent can do on its own with the gating controls that must be present.
- Suggest mode
- The agent proposes; a human chooses whether to apply. The default for new or high-risk integrations.
- Auto-apply mode
- The agent executes inside a sandboxed boundary (e.g., a PR branch) where a human still reviews before anything ships.
- Autonomous mode
- The agent both executes and ships. Reserved for low-risk, reversible, well-evaluated actions with strong monitoring.
- Required guardrail
- A control that *must* be in place at a given autonomy level — e.g., an audit log at every level, a content filter for any externally-visible output.
Common pitfalls
- Picking one autonomy level for the whole agent. Real systems have *per-action* autonomy — the same agent can be autonomous for `git status` and require approval for `npm publish`.
- Raising autonomy without raising the evaluation and monitoring budget. Higher autonomy means more reliance on signals; you need more evals, not fewer.
- Treating Suggest mode as a permanent solution. If everything stays in Suggest, the agent is providing no leverage; you are just paying for autocomplete.