Received None or empty response from LLM call when using Claude 3.5 Sonnet

[javadrip]

I was using 4o-mini to build and test, and when everything was running well (as in the code runs from start to finish, not that the quality of the output was fantastic), I switched to Claude 3.5 Sonnet. That was when I got the "Received None or empty response from LLM call".

Anyone else encountering this problem?

I tried the solution outlined here: https://community.crewai.com/t/why-am-i-getting-the-invalid-response-from-llm-call-none-or-empty-error-with-my-custom-tool-if-using-anthropic-llm-but-not-with-openai-llm/1571/8

I have been experiencing the same issue as well using the newest version of Claude Sonnet 3.5 through Amazon Bedrock, so it didn’t seem like it was related to model capability.

The only thing I have found to work is setting Agent(..., use_system_prompt=False). That would be my first recommendation since you are also using Anthropic.

The system prompt does not seem to be playing nice with Anthropic but I am not sure why.

It did work for me, but there's no answer from the CrewAI team about why or how this happened, and I'm not too comfortable with resorting to use_system_prompt=False in production when I don't know what system prompts are being removed.

Can someone from the CrewAI please address this?

[scottschluer]

Same issue happening occasionally with Claude 3.7 Sonnet when using SerperDevTool.

[caleblawrence]

same issue with gemini

[HelloWorldGitHubUser]

same issue happending when using gemini, but works well with openai/deepseek. Also it seems that there's hallucination/parsing problem using gemini/claude, where agents may tend to make up something according to the tool output.
Would crewAI team help address these, please? @lucasgomide @lorenzejay

[lucasgomide]

Hi folks! Does anyone have a Crew snippet?

The models most often mentioned are medium/large ones, so this issue shouldn't be happening. I also ran extensive tests using Claude 3.7 and Gemini models, and I didn't encounter any issues.

[HelloWorldGitHubUser]

Hi @lucasgomide
Here's a code snippet for my case:

# Task 1
task_1 = crew_setup.task_1()

crew_1 = Crew(
    agents=[crew_setup.agent_1()],
    tasks=[task_1],
    process=Process.sequential,
    verbose=True
)

crew_output = crew_1.kickoff(inputs=inputs)
result_1 = crew_output.raw
result_1 = clean_json_output(result_1)

types = result_1.get("types", [])

task_list = []

# Task 2 workflow
# Get the original task description from the config
shared_task_description = crew_setup.tasks_config['task_2']['description']

# Now create a task dynamically with the distinct context for each sort_type
for sort_type in types:
    type_name = sort_type.get("type_name", "Unknown type")
    type_description = sort_type.get("description", "No description provided")

    print(f"Creating task for type: {type_name}")

    task_description = (
        f"You are analyzing the '{type_name}' type. The description of this type is: {type_description}\n\n"
        f"{shared_task_description}"
    )

    task = Task(
        description=task_description,
        expected_output=...
        agent=crew_setup.agent_2(),
    )
    task_list.append(task)


# Start a crew for each task to run independently
for i, task in enumerate(task_list):
    type_name = types[i].get("type_name", "Unknown type")
    print(f"\nRunning aggregation task for type: {type_name}")

    single_task_crew = Crew(
        agents=[crew_setup.agent_2()],
        tasks=[task],
        process=Process.sequential,
        verbose=True
    )

    task_result = single_task_crew.kickoff(inputs=inputs).raw
    cleaned_result = clean_json_output(task_result)