deep-agent-hitl

Deep Agent Human-in-the-Loop

Configure approval gates that pause the agent and wait for human approval before executing sensitive operations.

When to Use

• When the agent should ask before deleting files
• When deploys require human approval
• When sending emails/messages needs confirmation
• When any destructive operation needs a gate

Available Operations

1. Configure interrupt_on per tool
2. Handle approval/rejection flow
3. Resume after approval
4. Custom approval decisions

Multi-Step Workflow

Step 1: Configure Interrupts

from deepagents import create_deep_agent
from langgraph.checkpoint.memory import MemorySaver

agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[delete_file, read_file, send_email, deploy],
    interrupt_on={
        "delete_file": True,       # Always ask
        "read_file": False,         # Never ask
        "send_email": True,         # Always ask
        "deploy": True,             # Always ask
        "write_file": {             # Custom decisions
            "allowed_decisions": ["approve", "reject", "modify"]
        }
    },
    checkpointer=MemorySaver()  # REQUIRED for interrupts
)

Step 2: Run and Handle Interrupts

config = {"configurable": {"thread_id": "session-1"}}

# Agent runs until it hits an interrupt
result = agent.invoke(
    {"messages": [{"role": "user", "content": "Delete the old log files"}]},
    config=config
)

# Check if agent is waiting for approval
if result.get("__interrupt__"):
    interrupt = result["__interrupt__"]
    print(f"Agent wants to: {interrupt['tool']} with args: {interrupt['args']}")

    # Approve
    result = agent.invoke(
        {"messages": [{"role": "user", "content": "approved"}]},
        config=config
    )

    # Or reject
    # result = agent.invoke(
    #     {"messages": [{"role": "user", "content": "rejected, don't delete those"}]},
    #     config=config
    # )

Step 3: Build Approval UI (FastAPI)