Streaming and Tools does not work with Anthropic

[haranrk]

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

The same issue as #2284 but when using Anthropic Provider.

In the following code example, the openai_model works but when run with anthropic_model, the agent early returns.

Here's the output.

Streaming response:
11:32:04.545 agent run
11:32:04.545   chat claude-3-7-sonnet-latest
Logfire project URL: https://logfire-us.pydantic.dev/mithra/api
I'll calculate 10 + 5 for you and provide the answer in a creative way.Streaming all messages:
[
    ModelRequest(parts=[SystemPromptPart(content='You are a helpful assistant that can perform math calculations\nUser
will give you a math problem and you need to call calculator tool to solve it.\nAnswer with random text where each line
starts with a number and total lines must be equal to result of the operation.\n/no_think\n',
timestamp=datetime.datetime(2025, 8, 12, 18, 32, 4, 545757, tzinfo=datetime.timezone.utc)), UserPromptPart(content='What
is 10 + 5?', timestamp=datetime.datetime(2025, 8, 12, 18, 32, 4, 545761, tzinfo=datetime.timezone.utc))]),
    ModelResponse(parts=[TextPart(content="I'll calculate 10 + 5 for you and provide the answer in a creative way."),
ToolCallPart(tool_name='calculator', args='{"a": 10, "b": 5, "operation": "add"}',
tool_call_id='toolu_015WCNrb2mPDjTthtiDp7fm2')], usage=Usage(request_tokens=477, response_tokens=110, total_tokens=587,
details={'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 477, 'output_tokens': 110}),
model_name='claude-3-7-sonnet-latest', timestamp=datetime.datetime(2025, 8, 12, 18, 32, 5, 500711,
tzinfo=datetime.timezone.utc)),
    ModelRequest(parts=[ToolReturnPart(tool_name='calculator', content='Tool not executed - a final result was already
processed.', tool_call_id='toolu_015WCNrb2mPDjTthtiDp7fm2', timestamp=datetime.datetime(2025, 8, 12, 18, 32, 6, 560609,
tzinfo=datetime.timezone.utc))])
]

Example Code

import asyncio
import random
from typing import Literal

import logfire
from pydantic_ai import Agent
from pydantic_ai.models.anthropic import AnthropicModel
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.anthropic import AnthropicProvider
from pydantic_ai.providers.openai import OpenAIProvider
from rich import print

from app.config import settings

logfire.configure()
logfire.instrument_pydantic_ai()


def roll_dice() -> str:
    """Roll a six-sided die and return the result."""
    return str(random.randint(1, 6))

# --- Agent and Model Configuration ---
# Replace with your api key
anthropic_model = AnthropicModel(
    "claude-3-7-sonnet-latest",
    provider=AnthropicProvider(api_key=settings.anthropic_api_key),
)

openai_model = OpenAIModel(
    "gpt-4o-mini",
    provider=OpenAIProvider(api_key=settings.openai_api_key),
)

default_system_prompt = """
You are a helpful assistant. 
Do not ask any follow-up questions.
"""

model_settings = {
    "temperature": 0.2,
}

tools = [roll_dice]
SYS_PROMPT = """You are a helpful assistant that can perform math calculations
User will give you a math problem and you need to call calculator tool to solve it.
Answer with random text where each line starts with a number and total lines must be equal to result of the operation.
/no_think
"""


agent = Agent(model=anthropic_model, system_prompt=SYS_PROMPT)

@agent.tool
def calculator(ctx, a: int, b: int, operation: Literal['add', 'subtract', 'multiply', 'divide']) -> int:
    print(f" > tool calculator: {a} {operation} {b}")
    if operation == 'add':
        return a + b
    elif operation == 'subtract':
        return a - b
    elif operation == 'multiply':
        return a * b
    elif operation == 'divide':
        return a // b  # Integer division
    else:
        raise ValueError("Invalid operation. Use 'add', 'subtract', 'multiply', or 'divide'.")


async def main():
    prompt = 'What is 10 + 5?'

    print('Streaming response:')
    async with agent.run_stream(prompt) as result:
        async for message in result.stream_text(delta=True):
            print(message, end='', flush=True)

        print('Streaming all messages:')
        print(result.all_messages())

    print()

    print('---' * 10)

    print('NON streaming response:')

    response = await agent.run(prompt)
    print(response.output)

    print('NON stream All messages:')
    print(response.all_messages())


if __name__ == '__main__':

    asyncio.run(main())

Python, Pydantic AI & LLM client version

0.6.2

[DouweM]

@haranrk run_stream is a bit inflexible in that it will run the agent loop silently until it sees something that looks like final output that you'd want to stream (like text, if you've not specified output_type or if str is included as an option), even if tool calls may still follow and this wasn't the "real" final output, as it has no way of knowing that at that point, and waiting to stream until we've received the entire response would defeat the point.

Continuing the run with tool calls and all after text has already been streamed is also problematic, because run_stream/stream_text has no way of communicating that you should discard the originally streamed "final result" because the run was not actually done yet.

If you want the ability to stream and make sure the entire agent run is completed including tool calls, we recommend using agent.iter() (https://ai.pydantic.dev/agents/#streaming), or the new event_stream_handler (#1977) that I expect to be released today or tomorrow (https://github.com/pydantic/pydantic-ai/blob/main/docs/agents.md#streaming-all-events).

Would either of those work for your use case?

Note that when using agent.iter(), the request_stream is an AgentStream that also has stream_text and stream_output methods:

async with agent.iter(...) as run:
    async for node in run:
    if Agent.is_model_request_node(node):
        async with node.stream(run.ctx) as request_stream:

[haranrk]

@DouweM thanks for the fast response, I will try that, and report back!

[glundgren93]

Im using the agent.iter() method and Im often getting early returns from tool execution as well.

[Request] Part 0 text delta: ' that for you.'
event PartStartEvent(index=1, part=ToolCallPart(tool_name='create_chart', tool_call_id='tooluse_rg2ZvGr1RhioKQGH-D4SQg'))
[Request] Starting part 1: ToolCallPart(tool_name='create_chart', tool_call_id='tooluse_rg2ZvGr1RhioKQGH-D4SQg')
event PartDeltaEvent(index=1, delta=ToolCallPartDelta(args_delta='', tool_call_id='tooluse_rg2ZvGr1RhioKQGH-D4SQg'))
[Request] Part 1 tool call delta: PartDeltaEvent(index=1, delta=ToolCallPartDelta(args_delta='', tool_call_id='tooluse_rg2ZvGr1RhioKQGH-D4SQg'))

[DouweM]

@glundgren93 I'm not quite sure what I'm looking at :) Can you please share your code, what you'd expect to see, and what you're seeing instead?