Human-in-the-Loop Design

The Autonomy Spectrum

Human-in-the-loop is not binary. It is a spectrum from full human control to full autonomy. The right point on this spectrum depends on the action's reversibility, the model's track record, and the time pressure.

When to Suggest Only

• Destructive actions — deleting resources, dropping databases, removing users.
• Actions with downstream impact — scaling services that affect billing or capacity.
• Novel situations — the model has not seen this exact scenario before.
• Ambiguous alerts — the model is not confident about the root cause.
• Policy-sensitive actions — changes to RBAC, network policies, or secrets.

When to Act

• Read-only queries — getting pod status, reading logs, checking metrics.
• Information gathering — searching runbooks, looking up past incidents.
• Notifications — posting status updates, alerting the team.
• Well-understood, reversible actions — restarting a specific pod, clearing a cache.

Keeping the Human Accountable

The human must always know they are responsible. The AI assists, but the human owns the decision. This means the AI must present information clearly and not make it too easy to blindly approve.

def present_action_for_approval(action, context):
    """Present an action with enough context for informed approval."""
    return f"""I recommend the following action:

**Command:** `{action['command']}`
**Risk level:** {action['risk_level']}
**Expected outcome:** {action['expected_outcome']}
**Side effects:** {action.get('side_effects', 'None identified')}

**Why this action:**
{action['reasoning']}

**Confidence:** {action['confidence']}

Type APPROVE to execute or CANCEL to abort.
Note: You are responsible for this action."""

Design Patterns