Human-in-the-Loop: Customize Agent Plan Interactively

This example demonstrates advanced usage of the smolagents library, specifically showing how to implement Human-in-the-Loop strategies to:

1. Interrupt agent execution after plan creation using step callbacks
2. Allow user interaction to review and modify plans (Human-in-the-Loop)
3. Resume execution while preserving agent memory
4. Modify plans in real-time based on user feedback, keeping the human in control

Human-in-the-Loop Key Features

Interactive Plan Review

• The agent creates a plan and pauses execution
• Users can view the complete plan before execution begins
• Options to approve, modify, or cancel the plan

Plan Modification

• Users can edit the agent's plan in real-time
• Modified plans are applied to the agent's memory
• Execution continues with the updated plan

Memory Preservation

• Using reset=False preserves the agent's memory between runs
• Demonstrates how to build on previous interactions
• Shows memory state management across multiple executions
• Maintains transparency and control

Usage

Basic Usage

python plan_customization.py

Key Components

Step Callback Function

def interrupt_after_plan(memory_step, agent):
    if isinstance(memory_step, PlanningStep):
        # Display plan and get user input
        # Modify plan if requested
        # Continue or interrupt based on user choice

Agent Configuration

agent = CodeAgent(
    model=InferenceClientModel(),
    tools=[DuckDuckGoSearchTool()],
    planning_interval=5,  # Plan every 5 steps
    step_callbacks={PlanningStep: interrupt_after_plan},  # Register callback for PlanningStep
    max_steps=10,
    verbosity_level=1
)

Resuming Execution

# First run - may be interrupted
agent.run(task, reset=True)

# Resume with preserved memory
agent.run(task, reset=False)  # Keeps all previous steps

Example Human-in-the-Loop Workflow

1. Agent starts with a complex task
2. Planning step is created automatically
3. Execution pauses for human review - step callback triggers
4. Human-in-the-Loop:
   1. User reviews the plan in a formatted display
   2. User decides to approve, modify, or cancel the plan
   3. User modifies the plan (if requested) - user can edit the plan
5. Execution resumes with approved/modified plan
6. Memory preservation - all steps are maintained for future runs, maintaining transparency and control

Interactive Elements

Plan Display

============================================================
🤖 AGENT PLAN CREATED
============================================================
1. Search for recent AI developments
2. Analyze the top results
3. Summarize the 3 most significant breakthroughs
4. Include sources for each breakthrough
============================================================

User Choices

Choose an option:
1. Approve plan
2. Modify plan
3. Cancel
Your choice (1-3):

Plan Modification Interface

----------------------------------------
MODIFY PLAN
----------------------------------------
Current plan: [displays current plan]
----------------------------------------
Enter your modified plan (press Enter twice to finish):

Advanced Features

Memory State Inspection

The example shows how to inspect the agent's memory:

print(f"Current memory contains {len(agent.memory.steps)} steps:")
for i, step in enumerate(agent.memory.steps):
    step_type = type(step).__name__
    print(f"  {i+1}. {step_type}")

Error Handling

Proper error handling for:

• User cancellation
• Plan modification errors
• Resume execution failures

Requirements

• smolagents library
• DuckDuckGoSearchTool (included with smolagents)
• Access to InferenceClientModel (requires HuggingFace API token)

Educational Value

This example teaches:

• Step callback implementation for custom agent behavior
• Memory management in multi-step agents
• User interaction patterns in agentic systems
• Plan modification techniques for dynamic agent control
• Error handling in interactive agent systems

Perfect for understanding how to build interactive, user-controlled AI agents that can adapt their behavior based on human feedback.