Advanced Steps

Beyond the foundational steps, CARL ships step types for MCP, skills, multi-agent orchestration, and more. Each gets a full guide in its own section; this page is the map with the essentials.

MCP — call tools on an MCP server

MCPStepDescription + MCPStepConfig. Transports: stdio, http, sse.

from mmar_carl import MCPStepDescription
from mmar_carl.models.config import MCPStepConfig, MCPServerConfig

MCPStepDescription(
    number=1,
    title="Call MCP tool",
    config=MCPStepConfig(
        server=MCPServerConfig(server_name="my_server", command="python", args=["-m", "my_mcp_server"]),
        tool_name="search",
        argument_mapping={"query": "$history[-1]"},
    ),
)

→ Full guide: MCP overview.

MCP resource — read a resource into memory

MCPResourceStepDescription + MCPResourceStepConfig. Fetches read-only data (files, docs, schemas) and writes it to memory + history.

from mmar_carl import MCPResourceStepDescription
from mmar_carl.models.config import MCPResourceStepConfig, MCPServerConfig

MCPResourceStepDescription(
    number=1,
    title="Load API reference",
    config=MCPResourceStepConfig(
        server=MCPServerConfig(server_name="docs", transport="sse", url="http://docs/sse"),
        resource_uri="docs://api/reference.md",
        output_memory_key="api_docs",
    ),
)

Agent handoff — delegate to a sub-chain

AgentHandoffStepDescription runs a complete sub-chain with an isolated context; inputs are resolved from parent memory/history and the result is merged back via config.output_memory_key.

from mmar_carl import AgentHandoffStepDescription
from mmar_carl.models.config import 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",
    ),
)

The sub-chain’s full ReasoningResult is available at step_result.result_data["sub_result"].

The rest at a glance

Step	Class · config	What it does	Example
AgentSkill	`AgentSkillStepDescription` · `AgentSkillStepConfig`	Run an AgentSkill folder (LLM / SCRIPT / HYBRID / SUBAGENT / LLM_AGENT modes; sandboxed).	agent_skill_example.py
Supervisor	`SupervisorStepDescription` · `SupervisorStepConfig`	LLM routes the task to one of N registered specialist sub-chains.	supervisor_routing_example.py
Debate	`DebateStepDescription` · `DebateStepConfig`	Round-robin role-based debate, then a judge synthesises.	—
Parallel sampling	`ParallelSamplingStepDescription` · `ParallelSamplingStepConfig`	Sample N independent responses, then vote / LLM-judge the best.	llm_council_example.py
Human input	`HumanInputStepDescription` · `HumanInputStepConfig`	Pause for human input (in-process callable or webhook).	human_in_the_loop_example.py
Tool discovery	`ToolDiscoveryStepDescription` · `ToolDiscoveryStepConfig`	Discover + register tools at runtime from a module / callables / dict.	—
Evaluation	`EvaluationStepDescription` · `EvaluationStepConfig`	Inline quality gate on a prior step’s output; reacts via `on_fail` (continue / abort / retry-with-feedback).	—