The Rise of Agentic AI: Google’s Opal Update Signals a Shift in Enterprise AI Strategy

For the past year, the enterprise AI community has debated the appropriate level of autonomy for AI agents. Too little autonomy results in expensive workflow automation, while too much can lead to unpredictable and potentially damaging outcomes, as seen with early adopters of tools like OpenClaw. This week, Google Labs released an update to Opal, its no-code visual agent builder, that offers a potential solution—and it carries lessons for every IT leader planning an agent strategy.

Adaptive Routing, Persistent Memory, and Human-in-the-Loop Orchestration

The update introduces an “agent step” that transforms Opal’s previously static workflows into dynamic, interactive experiences. Instead of manually specifying each model or tool and their order, builders can now define a goal and let the agent determine the best path to achieve it—selecting tools, triggering models like Gemini 3 Flash or Veo for video generation, and initiating conversations with users when more information is needed. This represents a significant shift towards more capable and adaptable agents.

Google’s update highlights three key capabilities that will define enterprise agents in 2026: adaptive routing, persistent memory, and human-in-the-loop orchestration. These capabilities are made possible by the rapidly improving reasoning abilities of frontier models like the Gemini 3 series.

The Shift from “Agents on Rails”

Early enterprise agent frameworks, such as initial versions of CrewAI and LangGraph, were constrained by the limitations of earlier models. These models weren’t reliable enough for open-ended decision-making, leading to “agents on rails”—tightly controlled workflows where every step was pre-defined by a human developer. While functional, this approach was limiting and couldn’t adapt to novel situations.

The Gemini 3 series, along with models like Anthropic’s Claude Opus 4.6 and Sonnet 4.6, represents a threshold where models have become reliable enough for planning, reasoning, and self-correction, allowing for the removal of some of these “rails.” Google’s Opal update acknowledges this shift, trusting the underlying model to evaluate goals, assess tools, and determine the optimal sequence of actions dynamically. This mirrors the viability of agentic workflows in Claude Code, now packaged into a consumer-grade product.

Memory: The Key to Production-Ready Agents

The second major addition in the Opal update is persistent memory, allowing agents to remember information across sessions—user preferences, prior interactions, and accumulated context. This creates agents that improve with leverage, rather than starting from scratch each time.

While the technical implementation of Opal’s memory system hasn’t been disclosed, the concept is well-established. Tools like OpenClaw use markdown and JSON files for memory, suitable for single-user systems. Enterprise deployments face the challenge of maintaining memory across multiple users and sessions while adhering to security and data retention policies. The distinction between single-user and multi-user memory is a critical, often overlooked, challenge.

Human-in-the-Loop: A Design Pattern, Not a Fallback

The third pillar of the Opal update is “interactive chat”—the ability for an agent to pause, ask for clarification, or present choices before proceeding. This is human-in-the-loop orchestration, and its inclusion in a consumer product is significant.

Effective agents in production aren’t fully autonomous; they recognize their limits and hand control back to humans when needed. This is a more reliable approach than fully autonomous systems. Opal’s approach is more fluid, with the agent deciding when human input is required based on the quality of information, rather than relying on pre-defined checkpoints.

Dynamic Routing with Natural Language

The final significant feature is dynamic routing, where builders can define multiple paths through a workflow and let the agent select the appropriate one based on custom criteria. Google’s example involves an executive briefing agent that adapts based on whether the user is meeting with a new or existing client.

This is similar to conditional branching in frameworks like LangGraph, but Opal’s implementation lowers the barrier by allowing builders to describe routing criteria in natural language, which the model then interprets.

An Agent Intelligence Layer

Google is building an intelligence layer that orchestrates complex tasks, recruiting models, invoking tools, managing memory, routing dynamically, and interacting with humans—all driven by the reasoning capabilities of Gemini models. This pattern is emerging across the industry, with Anthropic’s Claude Code utilizing similar principles.

Practical Steps for Enterprise Agent Builders

The foundational patterns for building effective AI agents are now productized. Enterprise teams can evaluate their current architectures, prioritize memory as a core component, design human-in-the-loop as a dynamic capability, and explore natural language routing to involve domain experts. While Opal itself may not become the standard enterprise platform, the design patterns it embodies—adaptive, memory-rich, human-aware agents powered by frontier models—will define the next generation of enterprise AI.