Ask the Agent a Question

Use a text-only stage when success depends on the agent answering a question.

from magma_core.simulation.stage import AskingBaseStage

stage = AskingBaseStage(
    question="Which button should be pressed next?",
    answer="sw1",
)

The answer becomes the semantic validation target. The template already represents an answer stage, so it does not need another completion-answer flag.

When the agent must inspect the scene first, set allow_tools_before_answer=True and list the allowed detection tool names. An empty allowlist means all tools when tool use is enabled. Make the budget appropriate by setting the stage's target_tool_calls and max_tool_calls consistently; the template's defaults are one each.

Questions depending on earlier rules belong in requests: derive the expected answer from the same pre-request state used to produce the question. Do not expose the answer as a new instruction to the agent.

The tool tester skips text-only stages. Validate response correctness with GEN and a semantic judge. See stage reference.