🤖 Understanding Agent Policy in Genie Framework

Quick Summary

Agent Policy is the brain 🧠 of the Genie framework that makes smart decisions about how to respond to users. Think of it as a very organized manager that always knows what to do next!

🎯 What is Agent Policy?

Agent Policy is a core component of the Genie framework that determines how the conversational agent should respond to user inputs. Unlike traditional LLM approaches that rely on the model to directly generate responses, Genie uses a symbolic module to compute agent responses in a more controlled and reliable way.

Key Difference

Traditional LLMs: "Let me generate whatever feels right" 🎲

Genie Agent Policy: "Let me follow these specific rules step by step" ✅

🤔 Why Use a Symbolic Agent Policy?

There are two main reasons:

Deterministic Responses 🎯
- By providing specific agent acts, the LLM generates responses that are governed by the policy
- No more random or unpredictable responses!
Better Instruction Following 📝
- LLMs often struggle with following complex developer-defined policies
- A symbolic module helps ensure all instructions are properly followed

⚙️ How Agent Policy Works

Input/Output Flow

Inputs 📥 - Previous dialogue state - Updated dialogue state from the parser - Worksheet specification

Outputs 📤 - New dialogue state - Set of agent acts to be performed

🔄 Core Steps of Agent Policy

graph TD
    A[Start] --> B[Handle KB Queries]
    B --> C[Process Field Assignments]
    C --> D[Execute Complete Tasks]
    D --> E[Request Missing Information]
    E --> F[End]

Handle Knowledge Base (KB) Queries 🔍

if new_kb_record:
    if missing_required_parameters:
        ask_user()
    else:
        execute_query()
        report_results()

Process Field Assignments ✍️

for field in new_assignments:
    if needs_confirmation:
        ask_user_to_confirm()
    else:
        execute_actions()
        update_state()

Execute Complete Tasks ✅

for task in complete_tasks:
    if not executed:
        execute_worksheet_actions()
        update_state()
        add_agent_acts()

Request Missing Information ❓

for field in required_fields:
    if is_empty(field):
        ask_user(field)
        break

🎭 Types of Agent Acts

The Five Core Acts

Report 📊
```
Report(query, answer)
```
- Reports query results to user
- Example: "Here are the restaurants I found..."
- Used for sharing KB query results
Confirm ✔️
```
AskForFieldConfirmation(worksheet, field, value)
```
- Verifies field values with user
- Example: "Just to confirm, you want to book for 7 PM?"
- Ensures data accuracy
Say 💬
```
Say(utterance)
```
- Delivers direct messages
- Example: "I'll help you book a restaurant."
- Used for general communication
Propose 💡
```
ProposeWorksheet(worksheet, parameters)
```
- Suggests new tasks/actions
- Example: "Would you like to make a reservation?"
- Includes initial parameter values
Ask ❓
```
AskField(worksheet, field, field_description)
```
- Requests specific information
- Example: "What date would you like to book for?"
- Includes field description

🎯 Agent Acts in Action

Restaurant Booking Example

# Initial Search
[
    Report(
        query="looking for restaurants in SF",
        answer=["Nice Place", "Good Food"]
    )
]

# Proposing Booking
[
    ProposeWorksheet(
        worksheet="BookRestaurant",
        parameters={"restaurant": "Nice Place"}
    )
]

# Gathering Details
[
    AskField(
        worksheet="booking",
        field="date",
        description="Preferred booking date"
    )
]

# Confirming Details
[
    AskForFieldConfirmation(
        worksheet="booking",
        field="time",
        value="19:00"
    )
]

🔄 Common Act Sequences

Typical Flows

Search & Book Flow 📝

# 1. Report search results
Report(query, results)
# 2. Propose booking
ProposeWorksheet("BookRestaurant", {...})
# 3. Ask for details
AskField("booking", "date")
# 4. Confirm final details
AskForFieldConfirmation("booking", "all")

Information Query Flow 🔍

# 1. Report answer
Report(query, answer)
# 2. Ask if more info needed
Say("Would you like to know anything else?")

Modification Flow ✏️

# 1. Confirm current value
AskForFieldConfirmation(worksheet, field)
# 2. Ask for new value
AskField(worksheet, field)
# 3. Confirm change
Report("update_status", result)

⚙️ Act Parameters

Parameter Details

Worksheet 📋
- The form/task being handled
- Examples: BookRestaurant, CourseEnroll
- Must be a valid worksheet type
Field 🏷️
- Specific data point needed
- Examples: date, time, guests
- Must exist in worksheet
Description 📝
- User-friendly field explanation
- Helps generate clear questions
- Should be descriptive
Value 💡
- Current or proposed value
- Used in confirmations
- Must match field type

🎨 Response Generation

Response Guidelines

Formatting 📝
- Be polite and friendly
- Use conversation context
- Keep responses natural

Priority ⚡

if agent_acts:
    # Always perform listed actions first
    execute_acts_in_order()
else:
    # Then handle direct responses
    respond_to_user()

Context 🔄
- Reference previous states
- Maintain conversation flow
- Use appropriate formality

Remember

Agent acts are the building blocks of conversation - they help create structured, reliable, and natural interactions! 🌟

🎬 Example Flow

Restaurant Booking Scenario

User: "I want to book a romantic restaurant in London" 🍽️

Agent Policy Process:
1. 🔍 Creates KB record to search restaurants
2. ✨ Checks required parameters (date, time needed)
3. ❓ Asks user: "What date would you like to book for?"
4. ⏰ After getting date, asks for time
5. 🎉 Once all parameters are present:
   - Executes restaurant search
   - Reports matching restaurants
   - Proposes booking task with found restaurant

🌟 Benefits of Genie's Agent Policy

Key Advantages

Reliable Context Management 🧠
- Never forgets important conversation details
Controlled Flow 🎮
- Ensures all required information is collected
Predictable Behavior 🎯
- Follows developer-defined rules consistently
Mixed Initiative Handling 🤝
- Handles both user and system initiatives effectively

📊 Performance Impact

Real User Study Results

Genie's agent policy achieved impressive results:

Metric	Score	vs GPT-4
Execution Accuracy	86.5%	+21.1%
Dialog Act Accuracy	89.2%	+20.1%
Goal Completion Rate	82.8%	+61.0%

🚀 Best Practices

Common Pitfalls to Avoid

Don't skip confirmation steps for critical actions
Always handle missing required fields
Keep track of conversation state
Handle user interruptions gracefully

🔍 Debugging Tips

Troubleshooting Guide

Check dialogue state consistency
Verify all required fields are being tracked
Monitor agent act sequences
Review KB query results
Validate field assignments