ReAct Planner + reasoning model = misplaced planning message

[bfinco]

Say we have a base agent using ReAct planner:

planner = PlanReActPlanner()

root_agent = Agent(
    name = "email_analyst",
    model = LiteLlm(model=AGENT_MODEL),
    instruction = instruction,
    tools = [],
    planner = planner,
)

Say we make a request when chatting with it.

When AGENT_MODEL is a non-reasoning model (gemini 2.0, gpt 4o, etc.), it will first lay out the plan in its first message, then call the tools, then reason, then provide the final answer.

But when AGENT_MODEL is a reasoning model (tested with OpenAI o3 and o4-mini), it will do all the tool calls first, and then at the end create one final message with the planning, reasoning and final answer (which might be split into a separate message occasionally).

Is this expected? Any adjustments to fix this so that it behaves as a non-reasoning model?

(I didn't find much documentation on the planners yet)