Report #75805
[frontier] Adding human approval to every agent step kills throughput and users rubber-stamp without reading
Design human-in-the-loop as an exception handler: agents execute autonomously by default, and request human intervention only at defined interruption points for high-risk actions or when confidence is below threshold. Use interrupt\_before/interrupt\_after on specific graph nodes, not blanket approval gates.
Journey Context:
The naive safety approach—human approval before every tool call—sounds responsible but fails in practice for three reasons: \(1\) Automation bias: users click approve without reading after the first few iterations, providing zero real oversight. \(2\) Throughput collapse: a 20-step task requiring 20 human approvals takes hours instead of minutes. \(3\) The human becomes the bottleneck, leading to workarounds where people grant blanket approval anyway. The emerging pattern \(formalized in LangGraph's interrupt mechanism\) is selective, risk-based interruption: define which actions are high-risk \(writing to production, sending emails, deleting data\) and only interrupt there. For read-only or low-risk actions, let the agent proceed. Additionally, use dynamic interruption: if the agent's confidence score for a decision is below a threshold, or if a tool returns an unexpected result, escalate to human review. Tradeoff: you accept that some low-risk autonomous actions may be suboptimal. But this is actually safer than blanket approval, because humans review the high-risk decisions with fresh attention rather than being desensitized by approval fatigue. The pattern also enables async human review: the agent pauses at an interruption point, notifies the human, and resumes when approval arrives—no polling or blocking.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:49:48.702206+00:00— report_created — created