Google’s Opal platform received a notable update this week that offers enterprise teams a clearer picture of where AI agent design is heading. The update, released through Google Labs, introduces what the company calls an “agent step” — a capability that shifts Opal from a static drag-and-drop workflow builder into something that can reason about goals, select its own tools, and initiate conversations with users when it needs more information.
The practical change is significant. Previously, builders using Opal had to pre-define every decision point: which model to call, in what order, under what conditions. The new agent step removes that requirement. A builder now defines a goal, makes tools available — including models like Gemini 3 Flash or Veo for video generation — and lets the underlying model determine the path forward dynamically.
This matters because it addresses a fundamental tension that has shaped enterprise agent development for the past two years. Early agent frameworks required what practitioners came to call “agents on rails” — tightly scripted workflows where every branching path had to be anticipated in advance. The approach was stable, but brittle. Any situation the developer hadn’t foreseen would break the system. And the anticipation problem compounds quickly: for anything beyond a simple linear task, mapping every possible state becomes genuinely unmanageable.
The rails existed for a reason. Earlier models weren’t reliable enough to handle open-ended planning without close constraint. That’s changing. The Gemini 3 series, alongside recent releases from Anthropic, represents a point where frontier models have become sufficiently capable at planning, self-correction, and reasoning that giving agents more autonomy is no longer the obvious liability it once was. Google’s own Opal update is a direct acknowledgment of that shift — packaged, notably, in a no-code consumer product rather than a developer-facing API.
For IT leaders still designing agent systems that pre-define every contingency, that packaging carries a message. The new design pattern inverts the old one: define goals and constraints, provide tools, and let the model handle routing. Less programming, more managing.
The second major addition is persistent memory. Opal agents can now retain information across sessions — user preferences, prior context, accumulated history — so that agents improve with use rather than resetting with each interaction. Google hasn’t disclosed the technical implementation, but the capability itself is well understood in the agent-building community and represents one of the clearest dividing lines between a polished demo and a production-ready system. An agent that forgets every conversation is an agent most enterprise users will abandon quickly.
Taken together, the Opal update sketches three capabilities that are likely to define enterprise agent architecture through the rest of 2026: adaptive routing, where the model selects its own path; persistent memory, where context accumulates over time; and human-in-the-loop orchestration, where the agent knows when to pause and ask rather than proceed blindly.
None of these ideas are new in isolation. What’s new is seeing them packaged into a no-code platform aimed at builders without deep engineering backgrounds. That’s a signal about maturity. Technologies move into no-code tooling when the underlying patterns have stabilized enough to be abstracted. Google appears to be betting that moment has arrived for agentic AI.
Photo by Team Nocoloco on Unsplash
This article is a curated summary based on third-party sources. Source: Read the original article
