earned-autonomy
Installation
SKILL.md
Produces a level per task, the evidence needed to raise it, and the triggers that lower it without asking.
The one-line law
Autonomy is granted per task, not per agent, and only against logged evidence. New tasks start at full review.
One agent can be trusted to qualify accounts unattended while its enrichment is still spot-checked. Granting autonomy to an agent rather than a task is how a good record on the easy work buys freedom on the risky work.
The four levels
| Level | What ships | What reaches the human |
|---|---|---|
| L0 Draft | nothing with wording in it | every worded item, before it goes |
| L1 Spot-check | items that passed the quality gate | a random sample, plus every low-confidence item |
| L2 Exception-only | items that passed the gate | flagged, edge and high-risk items only |
| L3 Auto | runs on schedule, ships on gate pass | a periodic digest, plus any guardrail trip |
The distinction that keeps L0 honest is copy risk versus click risk. Wordless actions carrying no claim, a connection request with no note, a profile visit, can move at L0 inside their rate limits, because nothing in them can misstate, leak, or embarrass. Anything with words waits.