Trust rules

Trust rules are the single place in a workspace that names which proposed actions may execute without a human looking first, and how confident the system has to be before that happens. There is one file per workspace, not one rule scattered per agent.

What a rule looks like

Each entry in trust.yaml names a registered action id (gmail.send, hubspot.update, and others defined in packages/core/src/libs/actions/), an autoApproveAbove confidence value between 0 and 1, and an enabled flag that defaults to false. A pending proposal for that action at or above the threshold executes automatically; everything else waits in the review queue.

rules:
  - action: hubspot.update
    autoApproveAbove: 0.95
    enabled: true
  - action: gmail.send
    autoApproveAbove: 0.99
    enabled: false

Why it matters in production

Without a rule set like this, every agent action either always pauses for a human — which does not scale past a handful of agents — or the code that calls an external system decides for itself when to skip review, which is hard to audit and easy to get wrong under load. Keeping the list short and the thresholds high, as the reference docs recommend, means most actions still land in a human's queue and the ones that don't are still logged.

How Vocion implements it

trust.yaml at the workspace root is validated against TrustManifestSchema (packages/core/src/libs/workspace/schemas.ts) and applied to the trust_rule table; the auto-approval check runs inside ActionService. Full reference: docs/entities/trust.md. Anything auto-executed this way still appears in the auto-executed list at /dashboard/review, so approval is audited even when it isn't asked for.

Related

• Glossary reference
• Deploying AI agents in production with human approval