For the past year, the enterprise AI community has been locked in a debate about how much freedom to give AI agents. Too little, and you get expensive workflow automation that barely justifies the “agent” label. Too much, and you get the kind of data-wiping disasters that plagued early adopters of tools like OpenClaw. This week, Google Labs released an update to Opal, its no-code visual agent builder, that quietly lands on an answer — and it carries lessons that every IT leader planning an agent strategy should study carefully.
The update introduces what Google calls an “agent step” that transforms Opal’s previously static, drag-and-drop workflows into dynamic, interactive experiences. Instead of manually specifying which model or tool to call and in what order, builders can now define a goal and let the agent determine the best path to reach it — selecting tools, triggering models like Gemini 3 Flash or Veo for video generation, and even initiating conversations with users when it needs more information.
It sounds like a modest product update. It is not. What Google has shipped is a working reference architecture for the three capabilities that will define enterprise agents in 2026:
-
Adaptive routing
-
Persistent memory
-
Human-in-the-loop orchestration
…and it’s all made possible by the rapidly improving reasoning abilities of frontier models like the Gemini 3 series.
The ‘off the rails’ inflection point: Why better models change everything about agent design
To understand why the Opal update matters, you need to understand a shift that has been building across the agent ecosystem for months.
The first wave of enterprise agent frameworks — tools like the early versions of CrewAI and the initial releases of LangGraph — were defined by a tension between autonomy and control. Early models simply were not reliable enough to be trusted with open-ended decision-making. The result was what practitioners began calling “agents on rails”: tightly constrained workflows where every decision point, every tool call, and every branching path had to be pre-defined by a human developer.
This approach worked, but it was limited. Building an agent on rails meant anticipating every possible state the system might encounter — a combinatorial nightmare for anything beyond simple, linear tasks. Worse, it meant that agents could not adapt to novel situations, the very capability that makes agentic AI valuable in the first place.
The Gemini 3 series, along with recent releases like Anthropic’s Claude Opus 4.6 and Sonnet 4.6, represents a threshold where models have become reliable enough at planning, reasoning, and self-correction that the rails can start coming off. Google’s own Opal update is an acknowledgment of this shift. The new agent step does not require builders to pre-define every path through a workflow. Instead, it trusts the underlying model to evaluate the user’s goal, assess available tools, and determine the optimal sequence of actions dynamically.
This is the same pattern that made Claude Code’s agentic workflows and tool calling viable: the models are good enough to decide the agent’s next step and often even to self-correct without a human manually re-prompting every error. The difference compared to Claude Code is that Google is now packaging this capability into a consumer-grade, no-code product — a strong signal that the underlying technology has matured past the experimental phase.
For enterprise teams, the implication is direct: if you are still designing agent architectures that require pre-defined paths for every contingency, you are likely over-engineering. The new generation of models supports a design pattern where you define goals and constraints, provide tools, and let the model handle routing — a shift from programming agents to managing them.
Memory across sessions: The feature that separates demos from production agents
The second major addition in the Opal update is persistent memory. Google now allows Opals to remember information across sessions — user preferences, prior interactions, accumulated context — making agents that improve with use rather than starting from zero each time.
Google has not disclosed the technical implementation behind Opal’s memory system. But the pattern itself is well-established in the agent-building community. Tools like OpenClaw handle memory primarily through markdown and JSON files, a simple approach that works well for single-user systems. Enterprise deployments face a harder problem: maintaining memory across multiple users, sessions, and security boundaries without leaking sensitive context between them.
This single-user versus multi-user memory divide is one of the most under-discussed challenges in enterprise agent deployment. A personal coding assistant that remembers your project structure is fundamentally different from a customer-facing agent that must maintain separate memory states for thousands of concurrent users while complying with data retention policies.
What the Opal update signals is that Google considers memory a core feature of agent architecture, not an optional add-on. For IT decision-makers evaluating agent platforms, this should inform procurement criteria. An agent framework without a clear memory strategy is a framework that will produce impressive demos but struggle in production, where the value of an agent compounds over repeated interactions with the same users and datasets.
Human-in-the-loop is not a fallback — it is a design pattern
The third pillar of the Opal update is what Google calls “interactive chat” — the ability for an agent to pause execution, ask the user a follow-up question, gather missing information, or present choices before proceeding. In agent architecture terminology, this is human-in-the-loop orchestration, and its inclusion in a consumer product is telling.
The most effective agents in production today are not fully autonomous. They are systems that know when they have reached the limits of their confidence and can gracefully hand control back to a human. This is the pattern that separates reliable enterprise agents from the kind of runaway autonomous systems that have generated cautionary tales across the industry.
In frameworks like LangGraph, human-in-the-loop has traditionally been implemented as an explicit node in the graph — a hard-coded checkpoint where execution pauses for human review. Opal’s approach is more fluid: the agent itself decides when it needs human input based on the quality and completeness of the information it has. This is a more natural interaction pattern and one that scales better, because it does not require the builder to predict in advance exactly where human intervention will be needed.
For enterprise architects, the lesson is that human-in-the-loop should not just be treated as a safety net bolted on after the agent is built. It should be a first-class capability of the agent framework itself — one that the model can invoke dynamically based on its own assessment of uncertainty.
Dynamic routing: Letting the model decide the path
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 is an executive briefing agent that takes different paths depending on whether the user is meeting with a new or existing client — searching the web for background information in one case, reviewing internal meeting notes in the other.
This is conceptually similar to the conditional branching that LangGraph and similar frameworks have supported for some time. But Opal’s implementation lowers the barrier dramatically by allowing builders to describe routing criteria in natural language rather than code. The model interprets the criteria and makes the routing decision, rather than requiring a developer to write explicit conditional logic.
The enterprise implication is significant. Dynamic routing powered by natural language criteria means that business analysts and domain experts — not just developers — can define complex agent behaviors. This shifts agent development from a purely engineering discipline to one where domain knowledge becomes the primary bottleneck, a change that could dramatically accelerate adoption across non-technical business units.
What Google is really building: An agent intelligence layer
Stepping back from individual features, the broader pattern in the Opal update is that Google is building an intelligence layer that sits between the user’s intent and the execution of complex, multi-step tasks. Building on lessons from an internal agent SDK called “Breadboard”, the agent step is not just another node in a workflow — it is an orchestration layer that can recruit models, invoke tools, manage memory, route dynamically, and interact with humans, all driven by the ever improving reasoning capabilities of the underlying Gemini models.
This is the same architectural pattern emerging across the industry. Anthropic’s Claude Code, with its ability to autonomously manage coding tasks overnight, relies on similar principles: a capable model, access to tools, persistent context, and feedback loops that allow self-correction. The Ralph Wiggum plugin formalized the insight that models can be pressed through their own failures to arrive at correct solutions — a brute-force version of the self-correction that Opal now packages some of that into a polished consumer experience.
For enterprise teams, the takeaway is that agent architecture is converging on a common set of primitives: goal-directed planning, tool use, persistent memory, dynamic routing, and human-in-the-loop orchestration. The differentiator will not be which primitives you implement, but how well you integrate them — and how effectively you leverage the improving capabilities of frontier models to reduce the amount of manual configuration required.
The practical playbook for enterprise agent builders
Google shipping these capabilities in a free, consumer-facing product sends a clear message: the foundational patterns for building effective AI agents are no longer cutting-edge research. They are productized. Enterprise teams that have been waiting for the technology to mature now have a reference implementation they can study, test, and learn from — at zero cost.
The practical steps are straightforward. First, evaluate whether your current agent architectures are over-constrained. If every decision point requires hard-coded logic, you are likely not leveraging the planning capabilities of current frontier models. Second, prioritize memory as a core architectural component, not an afterthought. Third, design human-in-the-loop as a dynamic capability the agent can invoke, rather than a fixed checkpoint in a workflow. And fourth, explore natural language routing as a way to bring domain experts into the agent design process.
Opal itself probably won’t become the platform enterprises adopt. But the design patterns it embodies — adaptive, memory-rich, human-aware agents powered by frontier models — are the patterns that will define the next generation of enterprise AI. Google has shown its hand. The question for IT leaders is whether they are paying attention.
Credit: Source link
