Description

When using the invoke() method of a CompiledStateGraph created by deepagents.create_deep_agent() with a non-None context parameter, Pydantic emits a serialization warning only when the agent invokes a subagent (calls the task tool). Direct responses without subagent invocation do not trigger this warning.

Environment

• deepagents: 0.2.7
• langgraph: 1.0.3
• pydantic: 2.12.4
• Python: 3.13

Steps to Reproduce

from deepagents import create_deep_agent
from dataclasses import dataclass
from typing import Optional
from langgraph._internal._typing import DataclassLike

@dataclass
class Context(DataclassLike):
    """Agent context configuration"""
    key: Optional[str] = None

# Create agent with context_schema
agent = create_deep_agent(
    model="gpt-4o",  # or any OpenAI model
    context_schema=Context,
    system_prompt="Conduct research and write a polished report.",
)

# Test 1: Simple greeting that gets answered directly (no warning)
print("Test 1: Direct response")
result1 = agent.invoke(
    {"messages": [{"role": "user", "content": "hi"}]},
    context=Context(key="12312")
)
print(f"Messages count: {len(result1['messages'])}\n")

# Test 2: Chinese greeting that triggers subagent call (warning appears!)
print("Test 2: Triggers subagent")
result2 = agent.invoke(
    {"messages": [{"role": "user", "content": "你好"}]},
    context=Context(key="12312")
)
print(f"Messages count: {len(result2['messages'])}")

Actual Behavior

Test 1 (Direct response): No warning, 2 messages in result

Test 1: Direct response
Messages count: 2

Test 2 (Subagent invocation): Warning appears, 4 messages in result (includes tool call)

Test 2: Triggers subagent
Messages count: 4
/path/to/pydantic/main.py:464: UserWarning: Pydantic serializer warnings:
  PydanticSerializationUnexpectedValue(Expected `none` - serialized value may not be as expected [field_name='context', input_value=Context(key='12312'), input_type=Context])
  return self.__pydantic_serializer__.to_python(

Expected Behavior

No warning should be emitted when passing a valid Context instance that matches the context_schema type, regardless of whether a subagent is invoked or not.

Analysis

The warning only appears when:

1. The context parameter is provided with a non-None value
2. The agent decides to invoke the task tool (spawning a subagent)

When the agent responds directly without invoking subagents, no warning occurs.

This suggests the issue is related to how state (including context) is serialized and passed during subagent invocation, specifically in the _return_command_with_state_update function in deepagents/middleware/subagents.py.

The warning message indicates Pydantic expects the context field to be None (based on the default value), even though Optional[Context] explicitly allows non-None values. This appears to be a type definition or serialization configuration issue when creating the internal state schema.

Impact

• Functionality: The code works correctly despite the warning
• User Experience: Confusing warning messages pollute logs
• Debugging: Makes it harder to identify real issues
• Production: Can cause concern in production environments with strict warning policies

Additional Notes

• The warning is non-deterministic from a user's perspective - it depends on the agent's internal decision to use subagents
• Simple queries that the agent can answer directly don't trigger the warning
• More complex queries or certain input patterns (like Chinese greetings) that cause subagent invocation will consistently trigger the warning

Related Issues

This may be related to other context handling issues in LangGraph 1.0, such as #6342 which discusses problems with using context and config together.