Steps Overview

A CARL chain is a list of steps. Each step is a typed description object that declares what to do and what it depends on. The DAG executor picks an executor based on the step’s type.

Typed API vs legacy

Use the typed classes (LLMStepDescription, ToolStepDescription, …) for new code — they give you type-checked fields and clear intent. The legacy unified StepDescription class is still accepted for backward compatibility; convert it with legacy_step.to_typed_step().

Fields every step shares

All step classes extend StepDescriptionBase and inherit these fields:

Field	Type	Default	Purpose
`number`	`int`	— (required)	Step number in the sequence.
`title`	`str`	— (required)	Human-readable title.
`dependencies`	`list[int]`	`[]`	Step numbers this step waits for.
`triggered_by`	`list[str]`	`[]`	Event names that gate this step (see event-driven steps). Becomes ready only when all listed events have been emitted and numeric `dependencies` are met.
`checkpoint`	`bool`	`False`	Mark as a RE-PLAN rollback checkpoint.
`checkpoint_name`	`str \| None`	`None`	Optional checkpoint label.
`replan_enabled`	`bool \| None`	`None`	Per-step RE-PLAN override (`None` = use chain policy).
`metrics`	`list[MetricBase]`	`[]`	Metrics evaluated after the step runs.
`loop_back_to`	`int \| None`	`None`	Step to loop back to — see loops.
`loop_config`	`LoopConfig \| None`	`None`	Loop condition + budget guard.
`cache`	`StepCache \| None`	`None`	Result memoization — see caching.

The step types

Foundational

Type	Class	What it does
`llm`	`LLMStepDescription`	Chain-of-thought reasoning with an LLM (the default).
`tool`	`ToolStepDescription`	Calls a registered Python function.
`memory`	`MemoryStepDescription`	read / write / append / delete / list on shared memory.
`transform`	`TransformStepDescription`	Data transforms with no LLM call.
`conditional`	`ConditionalStepDescription`	Branch to a step based on a condition.
`structured_output`	`StructuredOutputStepDescription`	LLM output constrained to a JSON schema.

Advanced

These are introduced on the advanced steps page and get full guides in their own sections (Orchestration, Skills, MCP, Evaluation).

Type	Class	What it does
`mcp`	`MCPStepDescription`	Call a tool on an MCP server.
`mcp_resource`	`MCPResourceStepDescription`	Read a named MCP resource into memory.
`agent_skill`	`AgentSkillStepDescription`	Run an AgentSkill folder.
`agent_handoff`	`AgentHandoffStepDescription`	Delegate to a complete sub-chain.
`supervisor`	`SupervisorStepDescription`	LLM routes the task to one of N sub-chains.
`debate`	`DebateStepDescription`	Round-robin multi-agent debate + judge.
`parallel_sampling`	`ParallelSamplingStepDescription`	Sample N responses, vote / judge the best.
`human_input`	`HumanInputStepDescription`	Pause for human input (callable or webhook).
`tool_discovery`	`ToolDiscoveryStepDescription`	Discover + register tools at runtime.
`evaluation`	`EvaluationStepDescription`	Inline quality gate on another step’s output.

Cross-cutting features

Loops — re-run a range of steps until a condition holds.
Caching — memoize a step’s result within a run.
Dynamic references — $history, $memory, … wire steps together.