Autonomy ladder
A five-level scale, set per mission, for how much an agent may do without a human — from draft-only to self-improving.
Last verified
The autonomy ladder is a five-level scale that sets how much a given agent mission may do without a human, from producing a draft no one has to act on up to changing its own approach. It is a property of the mission, not a global setting, so two missions on the same agent can sit at different rungs.
The five levels
Vocion's autonomyPolicy.level on a missions/<slug>.yaml manifest runs
1 to 5: 1 is draft only, 2 asks before any action, 3 acts within the rules a
human already set, 4 manages a goal on its own, 5 improves itself. Levels 1
and 2 gate every externally visible action; level 3 and above let actions run
unless the specific task flags approval. Internal analysis and drafting are
never gated, at any level — the ladder governs what leaves the system, not
what the agent is allowed to think about.
Why it matters in production
A single global "autonomous or not" switch does not survive contact with a real team of agents: a lead generation mission and a payroll mission need different defaults, and a team that starts at level 2 needs a path to level 3 without a code change. Setting the level per mission means that path is a one-line YAML edit that goes through the same review as everything else.
How Vocion implements it
autonomyPolicy.level is a field on the MissionManifestSchema
(packages/core/src/libs/workspace/schemas.ts), documented in
docs/entities/mission.md.
It sets the ceiling; trust.yaml still decides, action by action, whether a
specific proposal clears the confidence bar to auto-execute at level 3 or
above.