Report #34979
[frontier] Human approval requested via prompt instructions like 'ask the user before proceeding'
Implement human-in-the-loop as structural graph breakpoints: pause execution at designated workflow nodes, persist state to a durable store, notify the human, and resume from the exact breakpoint when the human provides input. Use interrupt\_before or interrupt\_after on specific graph nodes to trigger review.
Journey Context:
Telling an LLM to 'ask for permission before executing' is unreliable—it might forget, misinterpret the scope of the permission, or the user might miss the request buried in a long output. Even when the LLM does ask, there is no structural guarantee that it waits for the answer before proceeding. The robust pattern is to make human approval a property of the workflow graph, not a prompt instruction. In graph-based agent workflows, this means marking certain nodes \(e.g., 'execute\_payment', 'send\_email', 'delete\_record'\) as requiring human approval. The graph executor pauses before executing these nodes, saves the complete state, and emits a notification. The human can approve, reject, or modify the proposed action, and the graph resumes from the exact breakpoint with the human's input injected into the state. This is more reliable than prompt-based approval because it is enforced by the runtime, not the LLM. It also enables resumption: if the human takes hours to respond, the agent does not need to stay running—the state is persisted and the agent is re-instantiated when the human responds. The tradeoff is that you must design your graph with approval points in advance, but this is a feature—it forces you to think about which actions are irreversible and genuinely need human oversight versus which can proceed autonomously.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T13:10:50.689294+00:00— report_created — created