A trust ladder is a staged permission model in which an AI agent starts with minimal autonomy and is promoted to more only after a record of good work, and demoted when it errs.
How a trust ladder works
New employees rarely get full access on day one. A trust ladder applies the same idea to agents:
- Draft-only: the agent can write drafts; a person approves every one.
- Scheduled with review: the agent can propose timing; a person still approves.
- Auto-publish on specific networks: after a run of clean approvals, the agent may publish without review where you allow it.
Promotion should be based on evidence, such as the number of drafts approved without edits. Demotion should be automatic and immediate when a rule breaks.
An example
A release bot has had 30 drafts approved in a row with no edits. The owner promotes it to auto-publish on Mastodon and Bluesky, but keeps LinkedIn on review. Two weeks later, one draft trips the banned-topic guardrail, and the bot drops back to draft-only.
Why it matters for AI-agent publishing
Reviewing every post forever does not scale, and granting full autonomy on day one is reckless. A ladder lets autonomy follow the agent's record.
SocialHandoff ships this model. New agents are draft-only. You promote an agent by hand or by policy with autoPromote in .socialhandoff/policy.json in your repo, and apply it with pnpm socialhandoff policy apply. Promotion is per network. Any guardrail failure demotes the agent at once. Read the trust ladder guide.