Org autonomy profiles

Every org chooses how much machine autonomy to allow before HumanOS requires explicit human approval. The profile lives on cp_orgs.autonomy_profile and is loaded on each agent execution by @human/agent-runtime.

Overview

Profile	Behavior
paranoid	Every muscle invocation requires human approval — regardless of risk or reversibility.
balanced	Default. Critical risk always gated; high risk gated when irreversible; low/medium proceed when delegation allows.
aggressive	Only critical risk requires approval.

Runtime enforcement is in MuscleContextImpl (packages/agent-runtime/src/context/muscle-context-impl.ts), fed by loadOrgAutonomyProfileFromDb (packages/agent-runtime/src/org-autonomy-profile.ts).

Why this matters

• Sovereignty: humans set org-wide guardrails, not per-prompt vibes.
• Fourth Law: low confidence escalates separately — autonomy profile gates action, not uncertainty.
• Provenance: autonomy overrides are metered (agent.autonomy.override).

Set org profile (Control Plane)

Valid values: paranoid, balanced, aggressive. Missing or invalid DB values fall back to balanced.

Runtime loading (agent authors)

Agent runtime reads the profile once per execution context:

import { loadOrgAutonomyProfileFromDb } from '@human/agent-runtime';

// Called inside AgentRuntime.createExecutionContext()
const profile = await loadOrgAutonomyProfileFromDb(db, orgDid);
// → 'paranoid' | 'balanced' | 'aggressive'

Approval gating logic (simplified):

// packages/agent-runtime/src/context/muscle-context-impl.ts
function requiresHumanApproval(profile: AutonomyProfile, riskLevel: string, reversible: boolean) {
  switch (profile) {
    case 'paranoid':
      return true;
    case 'balanced':
      if (riskLevel === 'critical') return true;
      return riskLevel === 'high' && !reversible;
    case 'aggressive':
      return riskLevel === 'critical';
  }
}

When approval is required, muscles escalate via ctx.muscles → /v1/approvals or Command Plane escalations (see Human-in-the-Loop).

Per-agent override

Org profile is the default. Individual agents can override:

human api PUT "/v1/control-plane/agents/$AGENT_DID/autonomy" \
  --body '{"autonomy_level":"supervised"}'

Use org profile for policy; agent override for exceptions (e.g. read-only monitor agent on a paranoid org).

Use cases

• Finance org — paranoid during month-end close; relax to balanced afterward.
• Internal dev sandbox — aggressive for reversible low-risk connector calls.
• Enterprise default — balanced at org creation; audit autonomy ratio via Command Plane diagnostics.

Security considerations

See also

• Human-in-the-Loop
• Governance policy enforcement
• Community: HumanOS policy threshold and config
• API: PUT /v1/control-plane/orgs/:orgDid/autonomy