Multi-turn Conversations

class AgentState(TypedDict):
    user_message: str                    # Current user input
    conversation_history: List[dict]     # Previous messages
    actions: List[dict]                  # Planned actions
    results: dict                        # Tool execution results
    final_response: str                  # Agent's response
    # ... other fields

{
    "role": "user" | "assistant",
    "content": "message content",
    "timestamp": "2026-03-08T14:30:00Z"
}

# From agent/graph/nodes/plan_actions.py
for msg in conversation_history[-5:]:
    if msg.get("role") == "user":
        messages.append(HumanMessage(content=msg["content"]))
    elif msg.get("role") == "assistant":
        messages.append(AIMessage(content=msg["content"]))

User: "Add a meeting with John"
Agent: "I'd be happy to schedule a meeting with John. When would you like to meet?"

User: "Tomorrow at 2pm"
Agent: "How long should the meeting be?"

User: "One hour"
Agent: "Meeting with John scheduled for March 9th, 2:00 PM - 3:00 PM."

User: "Show me my calendar for today"
Agent: "You have 3 events today:
1. Team standup at 9am
2. Dentist appointment at 2pm
3. Project review at 4pm"

User: "Cancel the dentist appointment"
Agent: [Knows "the dentist appointment" refers to event #2]
"Are you sure you want to delete 'Dentist appointment' on March 8 at 2pm? Reply 'yes' to confirm."

User: "Yes"
Agent: "Event deleted successfully."

User: "Create a task to buy groceries"
Agent: "Task 'buy groceries' created successfully."

User: "Actually, make that due tomorrow"
Agent: [Updates most recently created task]
"Task updated with due date March 9, 2026."

User: "Add notes: milk, eggs, bread"
Agent: "Task notes updated."

User: "Show me emails from Sarah"
Agent: "Found 5 emails from Sarah:
1. 'Project Update' (March 8)
2. 'Meeting Notes' (March 7)
3. 'Budget Review' (March 6)
..."

User: "Read the first one"
Agent: [Retrieves full content of email #1]
"Subject: Project Update
From: Sarah Johnson
Date: March 8, 2026

The project is on track for Q2 launch..."

User: "Reply to it"
Agent: "I'm sorry, I can't send emails yet. Would you like me to draft a response you can copy?"

User: "What's on my calendar tomorrow?"
Agent: "You have 2 events tomorrow:
1. Team meeting at 9am
2. Lunch at noon"

User: "Add a task to prepare for the team meeting"
Agent: [Creates task related to event from previous query]
"Task 'Prepare for team meeting' created with due date March 8 (before your 9am meeting)."

User: "Also create an agenda document"
Agent: "I can create a task for that. Would you like me to add 'Create agenda document' to your task list?"

User: "Yes, due today"
Agent: "Task 'Create agenda document' created with due date March 8, 2026."

# Initial state
state = {
    "user_message": "Show me my calendar",
    "conversation_history": [],
    "actions": [],
    "results": {},
}

# After plan_actions node
state = {
    **state,
    "actions": [{"tool": "list_calendar_events", "args": {...}}],
}

# After execute_tools node
state = {
    **state,
    "results": {"a1": {"events": [...]}},
}

# After synthesize_response node
state = {
    **state,
    "final_response": "You have 3 events today: ...",
}

# After user message
conversation_history.append({
    "role": "user",
    "content": user_message,
    "timestamp": datetime.now().isoformat(),
})

# After agent response
conversation_history.append({
    "role": "assistant",
    "content": final_response,
    "timestamp": datetime.now().isoformat(),
})

# From plan_actions node
system_prompt = f"""
You are an AI assistant that helps users manage their calendar, email, and tasks.

Current date: {current_date}
Current time: {current_time}

You have access to three meta-tools for discovering and invoking actions.
"""

messages = [SystemMessage(content=system_prompt)]

# Add conversation history
for msg in conversation_history[-5:]:
    if msg["role"] == "user":
        messages.append(HumanMessage(content=msg["content"]))
    else:
        messages.append(AIMessage(content=msg["content"]))

# Add current message
messages.append(HumanMessage(content=user_message))

response = llm_with_tools.invoke(messages)

✓ "Cancel the dentist appointment"
✓ "Mark the first task as done"
✓ "Read it"
✓ "Change that to 3pm"

✗ "Cancel event_abc123"  # Technical IDs not needed

✓ Multi-turn:
   "Show my calendar"
   "Move the 2pm meeting to 3pm"
   "Add John as an attendee"

✗ Single turn:
   "Show my calendar and move the 2pm meeting to 3pm and add John"

User: "Cancel my meeting"
Agent: "Which meeting? You have:
1. Team standup at 9am
2. Dentist at 2pm"

User: "The second one"
Agent: "Dentist appointment at 2pm. Are you sure? Reply 'yes' to confirm."

if len(matching_items) > 1:
    return "I found multiple matches. Which one did you mean?"

if context_unclear:
    return "I'm not sure what you're referring to. Could you clarify?"

User: "Show me my calendar for the week"
# ... 10 more messages ...
User: "Delete the first event"  # Agent doesn't remember which event

User: "Show me my calendar for the week"
# ... 10 more messages ...
User: "Show me Monday's events again"
Agent: "Monday has: 1. Team meeting at 9am"
User: "Delete that"

User: "Show my tasks and my calendar"
Agent: [Lists both]
User: "Delete the first one"  # First task or first event?

User: "Delete the first task"
# OR
User: "Delete the first calendar event"

User: "Delete my dentist appointment"
Agent: "Are you sure? Reply 'yes' to confirm."
User: "Show me my tasks"  # Interrupts confirmation flow