Agent doesn't run when no user_prompt is provided. Instructions and message_history are ignored.

[restless]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When I run an agent, only providing message_history to it then it never calls the LLM and simply returns the last message from the provided message_history.

The official example from https://ai.pydantic.dev/message-history/#summarize-old-messages is affected.

So, running this:

user_prompt = None
summary = await summarize_agent.run(user_prompt, message_history=message_history)

• never calls the LLM
• returns the last message from the history as output
• new_messages() are an empty list

but this is fine:

user_prompt = ""
summary = await summarize_agent.run(user_prompt, message_history=message_history)

• calls the LLM
• returns summarized messages
• new_messages() includes ModelRequest (with "") and ModelResponse (with summary)

Example Code

from pydantic_ai import Agent
from pydantic_ai import ModelMessage
from pydantic_ai import ModelRequest
from pydantic_ai import ModelResponse
from pydantic_ai import TextPart
from pydantic_ai import UserPromptPart

summarize_agent = Agent(
    "openai:gpt-5-mini",
    instructions="""
Summarize this conversation to include all important facts about the user and
what their interactions were about.
""",
)

message_history = [
    ModelRequest(parts=[UserPromptPart(content="Hi, my name is James")]),
    ModelResponse(parts=[TextPart(content="Nice to meet you, James.")]),
    ModelRequest(parts=[UserPromptPart(content="I like cars")]),
    ModelResponse(parts=[TextPart(content="I like them too. Sport cars?")]),
    ModelRequest(parts=[UserPromptPart(content="No, cars in general.")]),
    ModelResponse(parts=[TextPart(content="Awesome. Which one do you like most?")]),
    ModelRequest(parts=[UserPromptPart(content="Fiat 126p")]),
    ModelResponse(parts=[TextPart(content="That's an old one, isn't it?")]),
    ModelRequest(parts=[UserPromptPart(content="Yes, it is. My parents had one.")]),
    ModelResponse(parts=[TextPart(content="Cool. Was it fast?")]),
]

user_prompt = None

summary = await summarize_agent.run(user_prompt, message_history=message_history)

print(summary.output)  # returns the last message from history: "Cool. Was it fast?"
print(summary.new_messages())  # returns []

Python, Pydantic AI & LLM client version

pydanticai: 1.21.0
LLM: gpt-5-mini

[DouweM]

@restless Thanks for the report, that looks like an unintended side-effect of #2562.

The responsible code is here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/_agent_graph.py

Lines 249 to 256 in d578b42

	elif isinstance(last_message, _messages.ModelResponse):
	if self.user_prompt is None:
	# Skip ModelRequestNode and go directly to CallToolsNode
	return CallToolsNode[DepsT, NodeRunEndT](last_message)
	elif last_message.tool_calls:
	raise exceptions.UserError(
	'Cannot provide a new user prompt when the message history contains unprocessed tool calls.'
	)

I suppose we should only go straight to the CallToolsNode if the final ModelResponse contains tool calls that need to be executed first. Let me look at a fix.