Human-in-the-Loop (in LangGraph)

Some actions are too consequential to let an agent take alone — sending the email, issuing the refund, deleting the records. Human-in-the-loop (HITL) lets a LangGraph agent pause mid-run, wait for a person to approve, edit, or reject, then resume exactly where it stopped. It's the difference between an agent you demo and one you'd let touch production.

💡 In one line: Human-in-the-loop pauses a LangGraph run at a checkpoint so a person can approve, edit, or reject before it continues.

Why HITL?

  • Risky actions — payments, deletion, external emails.
  • Trust and compliance — an audit trail of who approved what.
  • Quality — a human corrects the agent before the mistake ships.
  • Gradual autonomy — start supervised, loosen the gates as confidence grows.

How It Works: Interrupt & Resume

HITL rests on persistence. Because a checkpointer saves state at every step, LangGraph can stop, wait indefinitely (even across a restart), and then resume from the exact checkpoint once a human responds.

No checkpointer, no HITL — that's the hard requirement.

The interrupt Function

Call interrupt() inside a node to pause and surface a payload to the human:


 
Resume by invoking the graph with a Command:

 
The Approval Gate
 

      
        
        Whiteboard
      
      
    
  
 Whiteboard diagram  
 
 

HITL Patterns
 
     
  • Approve / reject — a gate before a risky tool call.
    •  
  • Edit state — the human fixes the agent's draft or arguments.
    •  
  • Review output — approve the final answer before it's sent.
    •  
  • Provide input — the agent asks a clarifying question.
    •  
  • Time travel — rewind to an earlier step and take a different path.
    •  
 
Static vs. Dynamic Interrupts
 
     
  • Static — configured at compile time (interrupt_before=["tools"]), always pausing at that node.
    •  
  • Dynamic — an interrupt() call inside a node, firing only when conditions warrant (e.g. only for refunds over ₹10,000).
    •  
 
Dynamic is usually better: pause only when it matters.
 
Durability
 
Because state is checkpointed, a paused run can wait minutes or days — through restarts and deploys — and still resume correctly. That's what makes HITL practical rather than a demo trick.
 
Best Practices
 
     
  • Gate only genuinely risky actions — approval fatigue is real.
    •  
  • Show the human enough context to decide (the full payload).
    •  
  • Use a durable checkpointer (Postgres) in production.
    •  
  • Log approvals for the audit trail.
    •  
  • Prefer dynamic interrupts over pausing on every step.
    •  
 
Summary
 
     
  • HITL pauses a run so a human can approve, edit, or reject before it continues.
    •  
  • It works via interrupt() + Command(resume=...), and requires a checkpointer.
    •  
  • Patterns: approve/reject, edit state, review output, ask for input, and time travel.
    •  
  • Dynamic interrupts (condition-based) beat static ones for most workflows.
    •  
  • Durable state means a run can wait days and still resume exactly where it stopped.Â