Handoff

AgentHandoffStepDescription runs a complete sub-chain inside the parent run. The sub-chain gets an isolated ReasoningContext derived from the parent, with inputs resolved from parent memory/history; its result is merged back via output_memory_key.

from mmar_carl import AgentHandoffStepDescription, AgentHandoffStepConfig

AgentHandoffStepDescription(
    number=3,
    title="Delegate to research agent",
    sub_chain=research_chain,
    config=AgentHandoffStepConfig(
        input_mapping={"input.topic": "$memory.input.topic"},
        output_memory_key="research_result",
    ),
)

AgentHandoffStepConfig

Field	Default	Purpose
input_mapping	{}	Map sub-chain memory keys ← parent references.
output_memory_key	—	Where the sub-chain’s final output is stored.
output_namespace	—	Namespace for the output.
propagate_failure	True	Whether sub-chain failure fails this step.
inherit_tools	—	Share the parent’s tools with the sub-chain.
timeout	None	Sub-chain timeout.

The sub-chain’s full ReasoningResult is always available at step_result.result_data["sub_result"]. sub_chain is runtime-only (not serialized).

Supervisor vs handoff

• Handoff — delegate to one fixed sub-chain.
• Supervisor — let an LLM choose among several.