Problem
Fully autonomous agents can make irreversible mistakes — deleting data, sending emails, making payments — with no opportunity for correction.
Solution
Design agent workflows with explicit human checkpoints that pause execution, surface the intended action and its consequences, and require human approval before proceeding.
Implementation (LangGraph pattern)
- Define "interrupt" nodes in the graph for sensitive actions (e.g., "send_email", "delete_records")
- When the agent reaches an interrupt node, serialise the current state and pause execution
- Notify the human (Slack message, UI notification, email) with: proposed action, reasoning, expected consequences
- On approval: resume from the checkpoint. On rejection: route to replanning node
- Set time limits — auto-reject if no human response within N minutes
When to interrupt
- Irreversible actions (send, delete, deploy, publish)
- High-cost operations (> $10 API call, > 1 hour compute)
- High-uncertainty steps (agent confidence < threshold)
Trade-offs
- Pro: Prevents catastrophic irreversible errors
- Pro: Builds user trust in agentic systems
- Con: Interrupt latency can be minutes to hours — not suitable for synchronous workflows
- Con: Human fatigue if too many interrupts are required (alert fatigue)
When to use
Customer-facing agents with real-world consequences; regulated industries.
When to avoid
Fully automated batch jobs with dry-run validation first.