Google's Opal just quietly showed enterprise teams the new blueprint for building AI agents

Google Labs' Opal update introduces "agent steps" — dynamic, goal-driven workflows that define three key capabilities for enterprise agents in 2026: adaptive routing, persistent memory, and human-in-the-loop orchestration.

The Shift: From "Agents on Rails" to True Autonomy

The first wave of enterprise agent frameworks tightly constrained every decision point. As models became more reliable, the "agents on rails" approach became a bottleneck — you couldn't scale without pre-defining every possible path.

Three Defining Capabilities

Adaptive routing: Instead of specifying which model or tool to call and in what order, builders define a goal and the agent determines the best path — selecting tools, triggering models (Gemini 3 Flash, Veo for video), and reasoning about the approach.

Persistent memory: Agents maintain context across interactions, enabling multi-step workflows that build on previous work.

Human-in-the-loop orchestration: The agent can initiate conversations with users when it needs more information, rather than failing or making assumptions.

Why It Matters

Opal is a no-code visual agent builder, making these patterns accessible to enterprise teams without deep technical expertise. It serves as a working reference architecture for how enterprise agents should be built.

The key insight: better reasoning models enable less constrained agent designs. When the model is reliable enough to make good decisions, you can remove the rails.

Source: VentureBeat