Skip to main content

Task Agent

Task automation with human approval gates.

Difficulty: Intermediate

What You'll Learn​

  • LangGraph interrupt() for human-in-the-loop
  • Resume from checkpoint after approval
  • Multi-step workflows
  • Durable execution

Workflow​

plan_task -> await_approval --(approved)--> execute_task -> END
                  |
                  +--(rejected)--> revise_plan --+
                                                 |
                  +------------------------------+

Setup​

agentvisor template create langgraph/task-agent
cd task-agent
docker compose up -d
agentvisor serve .

Test It​

Approval Flow​

THREAD=$(curl -sX POST http://localhost:8090/threads | jq -r '.thread_id')

# Step 1: Submit task (returns interrupted)
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=60s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"messages": [{"role": "user", "content": "Send an email to the team about the meeting"}]}}'

# Step 2: Approve
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=60s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"messages": [{"role": "user", "content": "approved"}]}}'

Rejection Flow​

# Reject with feedback
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=60s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"messages": [{"role": "user", "content": "rejected: make it more formal"}]}}'

Approval Commands​

CommandAction
approved, yes, okExecute the plan
rejected: <feedback>Revise with feedback
rejected, noRevise with generic feedback

Key Code​

Interrupt for Approval​

from langgraph.types import interrupt

def await_approval(state):
    interrupt({
        "status": "awaiting_approval",
        "plan": state.get("plan"),
        "message": "Reply 'approved' or 'rejected: <feedback>'"
    })
    return {}

Conditional Routing​

def check_approval(state) -> Literal["execute", "revise", "end"]:
    user_response = get_last_user_message(state)
    is_approved, feedback = parse_approval(user_response)

    if is_approved:
        return "execute"
    if state.get("approval_attempts", 0) >= 3:
        return "end"
    return "revise"

Durable Execution​

The workflow is durable:

  • State persisted in Temporal
  • Server can restart while awaiting approval
  • Resume continues exactly where it left off

Next Steps​

  • MCP Agent: Dynamic tool loading with MCP integration