Autonomy you grant one action type at a time

A level is set per action type, either for the whole workspace or for one cloud account. Every action type starts at approve each, and the level is enforced at the plan stage that every change passes through.

• Observe: nothing runs, even if a person approves. The run stops before any change and records why.
• Approve each: an owner or admin approves in the app, in Slack or through the API. Runs nobody approved are refused.
• Auto + rollback: a reversible action type may run without a person, and a failed post-check rolls the change back.

Levels combine the opposite way to safety policy: the most restrictive setting wins. A workspace level replaces the default, and an account level can lower it but never raise it.

An account doesn’t inherit a workspace-wide auto + rollback either. Without a level of its own it stays at approve each, because a record built in a sandbox account says nothing about production. If a level can’t be read, or a run’s account can’t be resolved, the run is treated as approve each.

Each finished run is classified from its stage log: for example a verified success, a success that couldn’t be verified, a block by a safety check or guardrail, a cloud error, or a rollback after a failed post-check. Only a verified success counts toward promotion, meaning the post-check had state to compare and it matched.

The record is kept per account and action type over a 30-day window, and it includes runs people approved. A single bad signal blocks promotion: a failed post-check that was rolled back, a rollback that failed, an irreversible change whose check failed, or a saving the ledger later found reverted or rolled back.

Settings → Autonomy lists every action type GetFinOps can run, for the workspace and for each AWS account, with its level, where that level was set, and progress toward the threshold: by default ten verified successes in 30 days with no bad signals. When an action type qualifies, Graduate raises it one level. The check runs again when you click, a recent demotion blocks it, and an account can’t graduate above the workspace’s level.

Dropping back needs no click. A post-check rollback, a saving found reverted at re-check, or a used safety override moves an action type from auto + rollback to approve each. The change is audited, you get an in-app notification, and the action type has to build a new record before it can graduate again.

After a scan, the plan’s safe actions are handed to the unattended lane. It can only run an action whose type is at auto + rollback for the account, and only when all of these hold:

• The workspace is on a paid plan or an active trial.
• The workspace has turned on agentic execution and automatic execution of safe actions, and has saved a guardrail configuration.
• The plan marked the action safe: every pre-check passed and no plan guardrail asked for a person.
• The workspace has fewer than ten queued, running or successful executions of that action type from the last 24 hours.
• The safety checks and guardrails pass again when the run starts.

• It doesn’t run anything on its own until the workspace turns on automatic execution and saves a guardrail configuration.
• It doesn’t promote anything by itself. Every step up is an owner’s or admin’s decision.
• It doesn’t let irreversible actions, such as deleting a volume or snapshot, run without a person.
• It doesn’t let an account’s level go above the workspace’s level. Raise the workspace level first.
• It can’t qualify every action type as safe yet. Some pre-checks aren’t measured, so stopping an EC2 instance, for example, always waits for a person today.
• It doesn’t apply to GCP, where GetFinOps makes no changes.