Descriptions:
Sam Witteveen examines the latest update to Google Labs’ Opal, a no-code visual agent builder that is transitioning from a rigid drag-and-drop workflow tool into a genuinely agentic system. The headline addition is a “generate step” — a node that lets the underlying model (Gemini) autonomously determine the best path through a workflow at runtime rather than following a pre-defined sequence. Witteveen connects this to a broader pattern he’s observed across frameworks: as models like Gemini 3 and Claude 4.5/4.6 improve at planning and reasoning, the scaffolding layer can afford to be less prescriptive.
Three new capabilities are dissected in detail: the generate step for dynamic path selection, dynamic routing for graph traversal (analogous to LangGraph patterns), and interactive chat for human-in-the-loop checkpoints where the agent pauses to request clarification before proceeding. Witteveen draws comparisons to OpenClaw — which became popular by autonomously building mini-workflows from natural language — positioning Opal as a safer, more consumer-accessible version of that paradigm.
A live demo shows Opal constructing a city events finder agent from a plain-language prompt, wiring together web search, maps, and weather tools with a structured output pipeline in under a minute. Witteveen argues this release signals that agent-building patterns previously confined to developer frameworks (CrewAI, LangGraph, OpenClaw) are now being productized for general users, and that Google is using Opal to learn which agent types people actually want before baking them into Gemini itself.
📺 Source: Sam Witteveen · Published February 27, 2026
🏷️ Format: Review
