Supervisor

SupervisorStepDescription asks an LLM to pick one specialist for the task, then runs that specialist’s sub-chain. Specialists are registered by name in the runtime-only agents field.

from mmar_carl import SupervisorStepDescription, SupervisorStepConfig

SupervisorStepDescription(
    number=1,
    title="Route to specialist",
    agents={"pdf": pdf_chain, "search": search_chain, "code": code_chain},
    config=SupervisorStepConfig(
        routing_prompt=(
            "Pick ONE specialist for the task. Reply with just the name.\n"
            "Specialists: {agents}\n\nTask: {task}"
        ),
        output_memory_key="specialist_result",
    ),
)

The routing_prompt can use {agents} (the available names) and {task} placeholders.

SupervisorStepConfig

Field	Purpose
routing_prompt	Prompt the LLM uses to choose a specialist.
task_source	Reference for the task text (default $history[-1]).
fallback_agent	Agent name to use if routing fails.
input_mapping	Inputs passed into the chosen sub-chain.
output_memory_key / output_namespace	Where the result is stored.
propagate_failure	Whether sub-chain failure fails this step.
inherit_tools	Share the parent’s registered tools with the sub-chain.
llm_config	Per-step model override for the routing call.

Note

agents is runtime-only (not serialized) — keyed by the names the router chooses from.

See also

• Handoff — delegate to a single fixed sub-chain.
• Supervisor routing example in the repo.