Autonomy and Human Roles

Risk-based execution modes, checkpoints, and human responsibilities

Autonomy classification and human operating roles

Rate ambiguity, reversibility, architectural consequence, familiarity, testability, data sensitivity, and external side effects low/medium/high. The most restrictive factor controls; no average cancels high sensitivity, irreversibility, or external effects.

ClassProfileHuman checkpointVerification / escalation
AutonomousLow consequence/ambiguity; familiar, reversible, strongly testable; no sensitive/external effectReview completionDeterministic evidence; escalate on scope change, check failure, or material assumption
SupervisedMedium factors; reversible with reliable reviewApprove boundary, review completionDeterministic plus targeted review; escalate on architecture/security impact or retry exhaustion
InteractiveHigh ambiguity/unfamiliarity, consequential architecture, or weakly testable synthesisApprove decisions as they ariseIndependent evidence/alternatives; escalate on conflict, low confidence, or boundary crossing
Human-ownedSensitive/legal judgment, irreversible action, credentials/production, consequential external effectHuman decides and executes/authorizesDual control/audit trail; AI never claims final authority

Human roles

  • Director sets outcome, scope, authority, risk appetite, checkpoints, and consequential choices.
  • Operator controls environments, credentials, external mutations, evidence, and rollback; confirms actual system state.
  • Reviewer checks traceability, evidence quality, correctness, safety, and acceptance criteria; rejects unsupported completion.

For research, the director defines the question, operator fixes/acquires sources, and reviewer checks citations/GRADE. For planning, the director prioritizes, operator mutates the tracker, and reviewer checks dependencies, testability, and architecture alignment.

Scenarios

1. Formatting a generated index with snapshot tests is autonomous. 2. Unfamiliar cross-service architecture is interactive with architecture-owner decisions. 3. Mixed-quality research synthesis is interactive because verification is inferential; require citation/GRADE review. 4. Publishing, deploying, purchasing, sending messages, or changing production access is human-owned unless narrowly authorized and reversible; the operator records evidence.