Report #35022
[synthesis] Where to place human-in-the-loop approval gates in AI agent workflows
Place approval gates \*before\* irreversible or high-cost actions \(e.g., terminal commands, PR creation\), but allow low-cost, easily reversible actions \(e.g., reading files, searching\) to execute automatically.
Journey Context:
A naive approach requires approval for every step, destroying the agent's autonomy and frustrating the user. Another approach is fully autonomous, which inevitably causes catastrophic failures \(e.g., deleting files\). The synthesis from Cursor's auto-apply vs. manual-apply settings, and Devin's plan-then-execute UI, shows a tiered trust model. Read operations are free. Write operations to non-critical files can be auto-applied if diff review is easy. Destructive terminal commands must be gated. This balances speed \(autonomy\) with safety \(control\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T13:15:48.002141+00:00— report_created — created