Python: InMemoryHistoryProvider auto-injection causes message duplication in HandoffBuilder workflows

[dantelmomsft]

Description

When agents participate in a HandoffBuilder workflow with CheckpointStorage, conversation history is managed by the HandoffAgentExecutor via its internal _full_conversation list. However, the Agent base class auto-injects an InMemoryHistoryProvider when context_providers is empty. This creates two independent history sources that compound on every turn, causing quadratic message growth and eventual API failures.

Reproduction

1. Create a HandoffBuilder workflow with CheckpointStorage and 2+ agents
2. Leave at least one agent with an empty context_providers list (the default)
3. Run a multi-turn conversation (3+ user messages)

Observed behavior

• Turn 1: Agent receives N messages (correct)
• Turn 2: Agent receives ~2N messages (executor history + InMemoryHistoryProvider replay)
• Turn 3: Agent receives ~3N messages
• Each turn inflates the message list further

This results in:

• The LLM re-requesting tool calls it already completed (duplicate task progress in the UI)
• OpenAI 400 errors: "An assistant message with 'tool_calls' must be followed by tool messages" — the duplicated history places tool_calls messages where the matching tool result is displaced
• Token usage growing quadratically instead of linearly

Expected behavior

Agents inside a HandoffBuilder workflow should not accumulate a second copy of conversation history. The executor is the single source of truth.

Current workaround

Create a no-op ContextProvider and assign it to every workflow participant that doesn't already have one:

class NoHistoryProvider(ContextProvider):
    def __init__(self):
        super().__init__("no_history_provider")

    async def before_run(self, *, agent, session, context, state):
        pass

Agent(
    client=client,
    instructions="...",
    name="my_agent",
    context_providers=[NoHistoryProvider()]  # prevents auto-injection
)

This is fragile — every new agent added to a workflow must remember to include it, and there's no framework-level indication that auto-injection is inappropriate.

For a real-world example of this workaround applied to a multi-agent banking assistant, see:
https://github.com/Azure-Samples/agent-openai-python-banking-assistant/blob/main/app/backend/app/agents/azure_chat/handoff_orchestrator.py#L57

Suggested resolution (pick one or combine)

Option	Description
A. Suppress auto-injection inside workflows	HandoffBuilder or HandoffAgentExecutor should disable InMemoryHistoryProvider auto-injection for all participants, since it owns the conversation state.
B. Add an explicit opt-out	Provide Agent(..., history_provider=None) or Agent(..., auto_history=False) so developers can explicitly disable it without a dummy provider.
C. Document the interaction	At minimum, add a warning in the HandoffBuilder / orchestration docs explaining that agents with empty context_providers will get InMemoryHistoryProvider auto-injected, which conflicts with the executor's history management. Include the no-op provider pattern as a recommended practice.

Environment

• agent-framework-core==1.0.0
• agent-framework-orchestrations==1.0.0b260402
• agent-framework-openai==1.0.0 / agent-framework-foundry==1.0.0
• Reproducible with both OpenAIChatCompletionClient and FoundryChatClient

[moonbox3]

Thanks for filing, @dantelmomsft. This is fixed in #5136.