Track Agent Steps, Monitor the Action

You're building a smolagents assistant for a basketball coach who needs help analyzing PDF reports that contain player stats, scouting insights, and game strategies.

The coach relies on the agent to answer questions like: "What defensive strategies should we run against their second unit?"

But the coach doesn't just want answers — they want visibility into what the agent is doing behind the scenes.

In this exercise, you'll write an action callback that runs every time the agent takes a step, such as calling a tool or using the model. This callback will:

Show the number of steps,
And if the agent has finished, display how many tokens were used.

This will help the coach (and you!) monitor how the agent is progressing and how much work it's doing to reach a conclusion.

本练习是课程的一部分

AI Agents with Hugging Face smolagents

练习说明

Complete the function signature by adding the agent_step parameter.
Check whether the current step produced a final answer using the .is_final_answer attribute of agent_step.
If it's the final answer, get the total number of tokens from total_tokens and print it.

交互式实操练习

通过完成这段示例代码来试试这个练习。

# Define an action callback that accepts the agent step and the agent
def action_callback(____, agent):
    step_num = agent_step.step_number
    print(f"Step {step_num}: Analyzing basketball data!")
    
    # Check if the agent step includes token usage
    if agent_step.____:
        total_tokens = agent_step.token_usage.total_tokens
        # Print how many tokens were used
        print(f"Analysis complete! Total tokens used: {____}")

编辑并运行代码

本练习是课程的一部分

AI Agents with Hugging Face smolagents

SkillTag.level.advancedSkillTag.label

4.8+

免费开始课程

Discover what makes code agents special and how they use Python to reason and act. Build your first agent with smolagents, add built-in and community tools for web access, and create custom tools to connect agents with data.

Exercise 1: Introduction to Hugging Face smolagents Exercise 2: Let the Agent Do the Math: Expense Insights Exercise 3: Wait, You Didn't Write Any Functions?Exercise 4: Creating an Agent With Tools Exercise 5: Writing Code Through Prompts: Compound Interest Exercise 6: Let Agents Search the Web Exercise 7: Creating an Agent With Custom Tools Exercise 8: Just Add @tool: Writing a Custom Tool Exercise 9: Check, Please! Giving Agents Access to Data Exercise 10: One Prompt, Many Tools

Transform your traditional RAG pipeline into an agentic system that retrieves information iteratively and reasons across multiple steps. Build stateful tools to support advanced retrieval, guide agents with planning intervals to improve outcomes, and use callbacks to track and customize agent behavior at runtime.

Exercise 1: RAG Refresher Exercise 2: RAG in Order: From Manuals to Answers Exercise 3: So, What’s Actually Happening in RAG?Exercise 4: The Step That Adds Meaning Exercise 5: Introduction to Agentic RAG Exercise 6: Stateful by Design: Build a RAG Search Tool Exercise 7: Connect the Tool, Unlock Agentic RAG Exercise 8: Working With Multi-Step Agents Exercise 9: Track Agent Steps, Monitor the Action

当前练习

Exercise 10: Game Time: Run the Full Agent

Tackle complex workflows by orchestrating teams of specialized agents under a coordinating manager. Add memory to retain context across interactions, debug agent behavior using execution traces and reasoning steps, and implement robust validation strategies to ensure high-quality, reliable responses.

Exercise 1: Introduction to Multi-Agent Systems Exercise 2: Divide and Conquer: Creating Specialist Agents Exercise 3: Meet the Team: Agent Coordination Exercise 4: Managing Agent Memory Exercise 5: Don't Forget: Keeping Memory Between Calls Exercise 6: Tracing the Agent’s Code Execution Exercise 7: Agent Final Answer Validation Exercise 8: Say More, Agent: Enforcing Answers Length Exercise 9: Smarter Agents: Validate and Retry Exercise 10: Congratulations!