Fix ValueError: agent_scratchpad should be a list of base messages, got str

How to Fix It

There are two working solutions reported in the sources. Both come from the accepted Stack Overflow answer by user furas, which references the LangChain GitHub discussion. The first solution is simpler and more aligned with the intended API. The second is a workaround that embeds the scratchpad directly into the system prompt string.

Solution 1: Replace MessagesPlaceholder with a string template for agent_scratchpad

This is the fix recommended by the accepted answer. Instead of using MessagesPlaceholder(variable_name="agent_scratchpad"), use a tuple with the role "ai" and the template "{agent_scratchpad}". This tells the prompt to expect a string variable named agent_scratchpad, which the agent executor will provide.

Step-by-step:

1. Locate the prompt definition in your code. It will look something like this:

from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder

prompt = ChatPromptTemplate.from_messages([
    ("system", """Respond to the human as helpfully and accurately as possible..."""),
    ("human", "{input}"),
    MessagesPlaceholder(variable_name="agent_scratchpad"),
])

2. Remove the MessagesPlaceholder line and replace it with a tuple that uses the "ai" role and a template string:

prompt = ChatPromptTemplate.from_messages([
    ("system", """Respond to the human as helpfully and accurately as possible..."""),
    ("human", "{input}"),
    ("ai", "{agent_scratchpad}"),
])

3. Keep the rest of your agent creation code unchanged. The agent executor will now pass the scratchpad as a string, which matches the template.

What to expect: The agent should create and run without the ValueError. The accepted answer confirms this works: the output shows ✅ Agent and Executor created successfully. and the agent executes correctly, producing a final result like {'input': 'call the tool', 'output': "The tool call was successful. The tool said: 'The tool says hello back!'"}.

When to use this: This is the primary fix. It works for the exact code pattern in the original question (using create_structured_chat_agent with FakeMessagesListChatModel). It is likely to work for any LangChain agent that uses create_structured_chat_agent.

Solution 2: Embed {agent_scratchpad} directly into the system prompt string

This is an alternative workaround also provided by the accepted answer. Instead of having separate message roles for human and agent_scratchpad, you can put both {input} and {agent_scratchpad} directly into the system prompt string as placeholders.

Step-by-step:

1. Modify your prompt so that the system message contains all the placeholders:

prompt = ChatPromptTemplate.from_messages([
    ("system", """Respond to the human as helpfully and accurately as possible. You have access to the following tools:

{tools}

Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).

Valid "action" values: "Final Answer" or {tool_names}

Provide only ONE action per $JSON_BLOB, as shown:

{{ "action": $TOOL_NAME, "action_input": $INPUT }}

Follow this format:

Question: input question to answer
Thought: consider previous and subsequent steps
Action:

$JSON_BLOB

Observation: action result
... (repeat Thought/Action/Observation as needed)
Thought: I know what to respond
Action:

{{ "action": "Final Answer", "action_input": "Final response to human" }}

Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation

{input}

{agent_scratchpad}
"""),
])

2. Remove the ("human", "{input}") and MessagesPlaceholder(variable_name="agent_scratchpad") lines entirely. The entire prompt is now a single system message.

What to expect: The accepted answer reports this also works without error. The agent executor will fill in both {input} and {agent_scratchpad} as strings into the system message.

When to use this: This is a fallback if Solution 1 does not work for your specific LangChain version or agent configuration. It is less conventional but equally functional.

If Nothing Works

If neither solution resolves the error, consider the following escalation paths based on the sources:

1. Check your LangChain version. The original question uses langchain==0.3.27, langchain-core==0.3.74, and langchain-community==0.3.27. If you are on a different version, the behavior of MessagesPlaceholder or create_structured_chat_agent may differ. Try upgrading or downgrading to match these versions.

2. Consult the LangChain GitHub issue tracker. The accepted answer references discussion #22885 in the langchain-ai/langchain repository. Search for that issue or open a new one with your exact code and version details.

3. Simplify your agent. As a temporary workaround, try using a simpler agent type like create_react_agent or create_tool_calling_agent which may not require agent_scratchpad in the same way. The structured chat agent is the one that triggers this error.

4. Use a different LLM. The original question uses FakeMessagesListChatModel for testing. If you are using a real model (e.g., from langchain-openai or langchain-aws), the error may still occur. The fix should be model-agnostic, but switching to a real model can help isolate whether the issue is with the fake model.

How to Prevent It

To avoid this error in future projects:

• Always use ("ai", "{agent_scratchpad}") instead of MessagesPlaceholder when working with create_structured_chat_agent. This is the pattern that works reliably based on the community reports. The MessagesPlaceholder approach is documented in some LangChain examples but appears to be incompatible with the structured chat agent in practice.

• Test with a fake model first. The original question uses FakeMessagesListChatModel from langchain_community.chat_models.fake. This allows you to catch prompt template errors like this one without incurring API costs or latency. Always validate your prompt template with a fake model before switching to a production LLM.

• Keep LangChain versions consistent. The error may be version-specific. If you upgrade LangChain, test your agent creation code immediately. The versions used in the working solution are langchain==0.3.27, langchain-community==0.3.27, langchain-core==0.3.74, langchain-aws==0.2.30, and langchain-openai==0.3.29.