LangChain / LangGraph - Darwin

Represent Darwin identifiers and states directly in your graph rather than keeping them only in messages. Copy the LangChain / LangGraph MCP recipe.

Suggested state

query → search results → selected capability revision → actionId → current Action state

Use separate nodes or tools for Search, Start, Get, Update, Approve, and Stop. Route from availableActions; do not infer valid transitions from elapsed time or model prose. Checkpoint before and after each Action mutation so retries can reuse the same requestId and avoid duplicate work.

Suggested nodes

Node	Reads	Writes
`search`	user intent and hard filters	canonical result page
`select`	result page and user choice	capability ID and revision
`start`	pinned revision, inputs, request ID	Action ID and current state
`route`	`availableActions` and status	next node
`interact`	exact Darwin interaction	user-controlled update or decision
`recover`	stored Action ID	refreshed current state

Never route directly from model sentiment such as “looks done.” Route from Darwin’s structured response and checkpoint the result.

Last modified on September 22, 2026