Most useful agents work alongside people rather than entirely alone. Designing the collaboration is as important as the automation.
Where People Add Value
- Clarifying goals and ambiguous requirements.
- Approving consequential actions.
- Reviewing outputs before they're used.
- Handling exceptions the agent can't resolve.
Design Patterns
- Plan review: approve the approach before execution.
- Action approval: confirm specific risky steps.
- Draft and review: the agent prepares, a person finalises.
- Escalation: the agent hands over when stuck or uncertain.
- Interruptibility: people can stop or redirect the agent at any time.
Make Review Easy
Present concise summaries, diffs and the key decisions, not pages of logs. Good review interfaces are what make oversight real.
Calibrate Autonomy
Start with more checkpoints; reduce them as the agent proves reliable on specific tasks. Keep checkpoints for irreversible or high-impact actions.
Avoid Rubber-Stamping
If people approve everything without reading, the checkpoint isn't working. Reduce low-value approvals so important ones get attention.