deepagents: collect each subagent's answer from the result messages, and test ProviderStrategy offline with a reactive fake model
deepagents 0.7.38, langchain 1.5.0, langchain-core 1.8.0. (1) To show what each subagent said without asking the lead to repeat it, walk result['messages'] after ainvoke: every AIMessage tool_call with name 'task' has args {'subagent_type','description'} and an id; the ToolMessage whose tool_call_id matches carries that subagent's final text. Map id -> subagent_type and you get one section per specialist, in the order asked, even when five task calls were issued in one AI message and ran in parallel. (2) response_format=ProviderStrategy(PydanticModel) works with a scripted BaseChatModel: make the lead's final AIMessage content a JSON string matching the schema and result['structured_response'] comes back as the Pydantic object. (3) For parallel subagents do not use a pop-from-a-list script (the order is not deterministic). Make the fake model reactive: in _generate, pick the role from the SystemMessage text, return tool calls if no ToolMessage is present yet, otherwise return the final text; have bind_tools return self.model_copy(update={'offered': names}) so each agent's offered tool names can be asserted separately. (4) create_deep_agent(tools=[]) still gives the lead the task tool; a subagent spec with its own 'tools' gets only those plus the filesystem tools and no task tool.