Secure "Human in the Loop" Interactions for AI Agents
summary
Argues that human oversight must sit between planning and execution for stateful AI agents and shows a working pattern using LangChain interrupts, persisted state and resume logic. Demonstrates a stock-trading demo that pauses to ask approval before buying 10 shares of SECO at $15, implements the pause by raising an InterruptException, and stores execution context (messages plus variables like destination and price) so the workflow persists state to Redis and resumes without re-running earlier steps. Travel-booking example shows why auto-execution without review fails.
AI-written summary · about
chapters (10)
- 0:00Secure human in the loop intro
- 1:09From chatbots to AI agents
- 3:37Why AI needs human oversight
- 7:11Risks of autonomous AI
- 8:39Human in the loop workflows
- 14:57Human review pauses AI
- 17:09Demo: human approval trading
- 18:35Code for human in the loop
- 21:21All for GenAI security
- 22:44Q&A: LangChain
related talks
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
27mBreaking the Chain: Agent Continuations for Resumable AI Workflows
43mAnnouncing New AI Capabilities
13mKeynote: Agentic AI Is Running Your Infrastructure
28mDesigning Empathetic, Resource-Aware AI Agents at Scale
49mAgentic AI Patterns
31mHow to Build Autonomous AI Agents That Run Anywhere
34mThe Human-in-the-Loop is Tired