06 of 6 · Requirement violation
The toggle
Autonomy granted because someone flipped a setting — not because the agent earned it on evidence.
In most stacks, "autonomy level" is a configuration value: someone chose it, once, in a settings page, and nothing about the agent’s actual track record can change it. The standard inverts this: autonomy is a graded privilege, earned per task-type on a published evidence trail, and revoked on one breach. The config version has a second defect — if configuration history can be overwritten in place, you can’t even prove who granted what, when.
Autonomy tiers stored as plain config with no linked evidence trail; elevation events with no graded track record behind them; no demotion path in code; config history that overwrites in place, so provenance can’t be reconstructed.
Earn it per task on evidence, revoke on one breach, and keep the highest-consequence actions human forever. Every elevation must trace to a graded record; every config write must append, never overwrite.
Part 7: promote-on-evidence / revoke-on-breach as substrate rules, with sticky locks for human-reserved actions; Part 8’s append-only constitution plane makes each grant’s provenance provable.
Official reference: Earned autonomy (Part 7) — related: AP-4 — the standard’s normative catalog defines ten anti-patterns; this page is the plain-English door into it.
The toggle — FAQ
Who decides the initial autonomy level, then?
A human — that never changes. What changes is what happens next: the level moves on the agent’s graded record, not on optimism. New task-type, new evidence trail, starting supervised. And the riskiest actions carry sticky locks no track record unlocks — earned autonomy has a ceiling by design.
Doesn’t revoke-on-one-breach make agents useless in practice?
It makes them careful in exactly the way you want: a breach demotes the agent for that task-type to supervised, where it keeps working — with review — while it re-earns the record. The alternative is what most stacks do today: nothing happens, because nothing is watching.
Does your fleet have this one?
The free Reality Check probes for this pattern against your own record — read-only, aggregate-only, no signup to read your result.
Run the Reality Check →