Approval fatigue begins when supervision becomes a ritual rather than a review. In one refactoring session, thirteen approval dialogs are enough to turn careful judgment into reflexive clicking; by the forty-seventh prompt, even a destructive action may receive an automatic Allow. More capable agents demand more oversight, but human attention remains finite, making supervision itself a throughput bottleneck. section 5.2.2 section 5.2.3
Reduce approval fatigue by moving routine decisions out of the attention-critical path while preserving friction for consequential actions:
- Auto-approve genuinely low-risk classes such as read-only access, directory listings, tests, and linting, while treating context-sensitive resources such as .env files cautiously. section 5.2.2
- Apply least privilege through project-specific tool allowlists; withhold database access or other capabilities that the task does not require. section 5.2.2
- Batch approvals where actions share a bounded risk profile, then retain close review for novel, destructive, or high-impact operations. section 5.2.2 section 5.2.3
The alternative is not simply removing humans from the loop. Headless agents work well with crisp inputs, bounded actions, and cheap validation, but unsupervised failures can remain unwitnessed until they become difficult to ignore. Use interactive supervision for novel or destructive work, and move repetitive, event-triggered tasks into terminal or CI automation only when validation and observability are strong. section 5.3 section 5.8
