Trust & human-in-the-loop

Where to add human checkpoints in agent workflows

The autonomy spectrum

Not all agent tasks should be fully autonomous. The right level of human involvement depends on the reversibility of actions, the stakes of getting it wrong, and how well-understood the task is.

The four oversight models

1. Fully supervised (human approves every step)

Best for: high-stakes irreversible actions, new workflows you’re still calibrating

2. Checkpoint model (human approves at key gates)

Best for: most agency workflows — efficient but controlled. Agent handles the volume, you approve at meaningful decision points.

3. Exception model (human only handles edge cases)

Best for: well-understood routine tasks with clear rules. Agent handles all triage, routes uncertain ones for human review.

4. Fully autonomous (no human in loop)

Best for: read-only or easily reversible tasks with very clear parameters. Example: agent compiles a daily briefing document.

Designing checkpoints

When building an agent workflow, ask these questions at each step:

• Is this action reversible? (If no → checkpoint before it)
• What’s the cost of getting this wrong? (High cost → checkpoint)
• Has this workflow run successfully 20+ times? (If not → checkpoint)
• Does this action affect external parties? (If yes → checkpoint)

Communicating trust to Claude

"You have permission to:
✓ Read my email and calendar
✓ Create drafts (do not send)
✓ Create files in Drive
✓ Search the web

You must stop and ask before:
✗ Sending any email or message
✗ Deleting anything
✗ Creating calendar events on behalf of others
✗ Making any purchase or financial commitment"