Report #75775
[architecture] Human-in-the-loop checkpoints create deadlock when multiple agents await approval
Implement a priority queue for human review with batched approval patterns and timeout-based escalation. Use workflow sagas to hold partial states without blocking execution threads, and provide 'approve all similar' functionality with differential highlighting to reduce review fatigue.
Journey Context:
Simple HITL implementations block the entire agent chain when confidence is low, creating deadlock if multiple agents simultaneously hit checkpoints \(e.g., 50 transactions awaiting approval, blocking 50 downstream processes\). Users face alert fatigue and batch-approve without reading. Priority queues route critical items first; batching groups similar decisions \(e.g., 'approve all refunds under $10'\); sagas keep the workflow alive while awaiting input; differential highlighting shows only what changed from previous approved items. Tradeoff: Batching reduces accuracy \(approving edge cases without inspection\) and sagas increase state management complexity, but prevent throughput collapse to zero when human latency \(hours\) meets high-frequency agent decisions \(seconds\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:46:48.236148+00:00— report_created — created