Generate from a Description

ChainBuilder.from_description is a meta-agent: it asks an LLM to plan a chain of LLM / Tool / Memory / Transform / Conditional steps for a task, then parses the plan through ReasoningChain.from_dict so all the usual validation applies (cycle detection, dependency references, reference-syntax warnings).

from mmar_carl import ChainBuilder

chain = await ChainBuilder.from_description(
    task="Outline the key arguments in the text, then condense them to 3 bullets.",
    llm_client=client,
    max_steps=4,
    max_retries=2,                       # self-correct on validation errors
    available_tools=["fetch", "summarise"],
)

Parameters

Parameter	Type	Default	Purpose
task	str	— (required)	Natural-language description of what the chain should do.
llm_client	Any	— (required)	Async client with get_response_with_retries(prompt, retries=…) or get_response(prompt).
available_skills	list[str] | None	None	Skill names surfaced in the planning prompt.
available_tools	list[str] | None	None	Tool names the planner may reference (tool steps must use these).
max_steps	int	10	Upper bound on planned steps (raises ValueError if exceeded).
max_workers	int | str	"auto"	Worker setting on the resulting chain.
extra_instructions	str	""	Free-form text appended to the planning prompt.
max_retries	int	2	Self-correction rounds when the produced chain fails validation.

It’s an async classmethod — await it.

Provenance

The full planner trail is written into chain.metadata for offline diagnosis: planner_prompt, planner_reply, and the per-attempt planner_attempts log (including validation errors that triggered a retry).

See also

• chain_from_description example in the repo.