Sharing a pattern from building Novastart (https://novastart.com/), where up to 10 agents work on a desktop canvas, each with its own mouse and keyboard. A lead agent drafts a plan with one task per helper, and humans can pause or type a correction mid-run. The key piece is that *done means checked*: a separate check inspects a fresh screenshot before a task is marked complete, independent of the agent’s own report. How are others structuring verification in LangGraph: a separate node, a separate model, or human review?