Depurando processos de raciocínio de agentes
Aplique abordagens sistemáticas para identificar e resolver problemas nas sequências de raciocínio e ação do seu agente.
Depurando processos de raciocínio de agentes é uma aula grátis de AI Agents with LangChain & Autonomous Workflows no CoddyKit. Esta é a aula 2 de 4. Você pode ler a aula completa abaixo gratuitamente — depois pratica ao vivo no navegador com um editor de código integrado e um tutor de IA 24/7. Faz parte do caminho de aprendizado de AI Agents with LangChain & Autonomous Workflows, e seu progresso é sincronizado entre a web e o app CoddyKit. O curso de AI Agents with LangChain & Autonomous Workflows inclui 4 aulas no total.
Partes desta aula ainda não foram traduzidas e aparecem em inglês.
Debugging Agent Thoughts
Ever had an AI agent give a weird answer or get stuck? Debugging agents isn't like debugging regular code. Instead of just finding syntax errors, we need to understand the agent's "thought process".
This lesson will teach you how to peek into your agent's mind to see why it makes certain decisions and how to fix its reasoning.
The Agent's Inner Voice
An AI agent doesn't just output a final answer. Internally, it goes through a series of "thoughts". These thoughts involve:
- Reasoning: What's the best next step?
- Tool Selection: Which tool should I use?
- Tool Input: What input should I give the tool?
- Observation: What was the result of using the tool?
By examining this sequence, we can pinpoint where the agent's logic might be failing.
Common Agent Issues
Agents can fail in several ways beyond simple code bugs:
- Wrong Tool: Selecting an irrelevant tool for the task.
- Bad Tool Input: Providing incorrect or malformed input to a tool.
- Reasoning Errors: Misinterpreting the problem or tool observations.
- Infinite Loops: Getting stuck in a repetitive cycle of thoughts and actions.
- Hallucinations: Making up facts or confidentially incorrect information.
Understanding these helps you know what to look for.
Seeing Agent Steps with Verbose
LangChain provides a simple way to see an agent's internal steps: the verbose=True parameter. When you set this, the agent will print its entire thought process to the console as it executes.
This "log" includes every Thought, Action, Action Input, and Observation, giving you a complete picture of its decision-making journey.
Tracing a Basic Agent
Let's see verbose=True in action. This agent uses a simple tool to get information. Pay attention to the output in the console!
Note: This code requires an OpenAI API key. Set OPENAI_API_KEY as an environment variable or uncomment and replace "YOUR_KEY".
import os
from langchain_openai import ChatOpenAI
from langchain.agents import AgentExecutor, create_react_agent, Tool
from langchain import hub # For standard prompts
# 1. Define a simple tool
def get_info(topic: str) -> str:
"""Provides info on simple topics."""
if "python" in topic.lower():
return "Python is a popular language."
elif "agent" in topic.lower():
return "An agent uses an LLM to decide actions."
return f"No specific info for '{topic}'."
tools = [
Tool(
name="info_tool",
func=get_info,
description="Useful for getting basic info on a topic.",
),
]
# 2. Set up the LLM (requires OPENAI_API_KEY)
# os.environ["OPENAI_API_KEY"] = "YOUR_KEY"
llm = ChatOpenAI(temperature=0, model="gpt-3.5-turbo")
# 3. Get the standard ReAct prompt
prompt = hub.pull("hwchase17/react")
# 4. Create the agent
agent = create_react_agent(llm, tools, prompt)
# 5. Create an agent executor with verbose logging
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)
# 6. Run the agent to see its thought process
agent_executor.invoke({"input": "What is an agent?"})Decoding Agent Logs
The verbose output shows a clear sequence:
> Entering new AgentExecutor chain...: Agent starts.Thought:: The LLM's reasoning for the next step.Action:: The name of the tool chosen.Action Input:: The arguments passed to the tool.Observation:: The result returned by the tool.Final Answer:: The agent's final response after its thoughts.
This structure is your roadmap for debugging!
Wrong Tool for the Job?
One common issue is the agent selecting the wrong tool or providing bad input. Look at the Action: and Action Input: lines.
- Did it pick a tool that doesn't fit the query?
- Did it extract the wrong information from the query to pass to the tool?
If so, you might need to refine your tool's description or adjust the agent's main prompt to guide it better.
Fixing Agent's Logic
If the agent's Thought: itself seems off, it's a reasoning problem. The LLM might be:
- Misunderstanding the overall goal.
- Failing to incorporate previous
Observations:. - Struggling with complex instructions.
To fix this, clarify the agent's system prompt, provide more context, or break down complex tasks into simpler sub-tasks.
Breaking the Loop
An agent stuck in an infinite loop will repeatedly generate similar Thought:, Action:, and Observation: sequences without progressing to a Final Answer:.
Common causes include:
- Ambiguous tool descriptions.
- Tools returning unhelpful or identical results.
- Prompts that don't clearly define a "completion" state.
Refine tool descriptions, ensure tools provide distinct outputs, or add explicit stopping conditions to your prompt.
Spot the Bug!
An agent is designed to summarize text. Here's a snippet of its verbose trace when asked to summarize "The quick brown fox jumps over the lazy dog":
Thought: I need to summarize the text.
Action: text_summarizer_tool
Action Input: What is the weather like?
Observation: The weather is sunny.
Thought: I need to summarize the text.
Action: text_summarizer_tool
Action Input: What is the weather like?
Observation: The weather is sunny.What is the primary debugging issue here?
Debugging Agents: Key Takeaways
Congratulations! You've learned how to systematically debug your AI agents. Key points:
- Use
verbose=Trueto expose the agent's internal thought process. - Examine Thought, Action, Action Input, and Observation.
- Identify issues with tool selection, tool input, or the LLM's reasoning.
- Address infinite loops by refining prompts, tool descriptions, or tool outputs.
Happy debugging, and build more robust agents!
Perguntas Frequentes
A aula “Depurando processos de raciocínio de agentes” é grátis?
Sim — o texto completo de “Depurando processos de raciocínio de agentes” é grátis para ler aqui na web. Para praticá-la interativamente (um editor de código integrado e um tutor de IA 24/7) e desbloquear o restante do curso de AI Agents with LangChain & Autonomous Workflows, atualize para CoddyKit PRO. O curso de AI Agents with LangChain & Autonomous Workflows inclui 4 aulas no total.
O que vou aprender em “Depurando processos de raciocínio de agentes”?
Aplique abordagens sistemáticas para identificar e resolver problemas nas sequências de raciocínio e ação do seu agente. Você pratica AI Agents with LangChain & Autonomous Workflows com código prático que executa diretamente no navegador, e um tutor de IA 24/7 responde suas dúvidas enquanto trabalha na aula.
Preciso ter experiência prévia para começar AI Agents with LangChain & Autonomous Workflows?
Nenhuma experiência prévia é necessária. AI Agents with LangChain & Autonomous Workflows no CoddyKit é estruturado para alunos iniciantes até avançados, então você pode começar aqui ou desde o início e aprender no seu ritmo. Esta é a aula 2 de 4.
Quanto tempo leva a aula “Depurando processos de raciocínio de agentes”?
A maioria das aulas CoddyKit leva cerca de 5–10 minutos. Cada uma é compacta e interativa, então você faz progresso constante e retoma exatamente de onde parou entre web e app.
Posso escrever e executar código nesta aula de AI Agents with LangChain & Autonomous Workflows?
Sim. Cada aula de AI Agents with LangChain & Autonomous Workflows inclui um editor de código integrado, então você escreve e executa código real direto no navegador e recebe feedback de IA instantaneamente — nenhuma configuração local necessária.
Todas as aulas deste curso
- LangSmith para rastreamento e monitoramento
- Depurando processos de raciocínio de agentes
- Avaliando o desempenho dos agentes
- Uso de tokens e monitoramento de custos