Google Opal turns workflow steps into self-directing agents

Building a custom AI workflow used to mean hardcoding every single step. You had to pick which model to call, guess every edge case, and pray the chain wouldn't break on a vague prompt.

Google Labs launched a major upgrade to Opal, its workflow-building app. The tool now includes a dedicated agent step available to all users worldwide.

Instead of forcing builders to manually select a model in the generate phase, Opal now lets users drop in an autonomous agent. The agent looks at the stated goal and automatically summons the right models and tools, like Web Search for gathering research or Veo for generating video clips.

According to reports, the new agent step runs on Google's Gemini 3 Flash model.

From static pipelines to dynamic teammates

Google illustrated the shift using interior design. In the past, an interior styling Opal was a one-way street: you uploaded a room photo, selected a style, and got a flat rendering back.

With the agent step, the app functions more like a creative collaborator. You upload an empty living room and ask for a mid-century look, and the agent delivers an initial concept. If something looks off, you give direct feedback, and the agent will even research niche design sub-styles online to refine the space.

The same dynamic applies to creative writing. Instead of rigid page counts, an agent-led storytelling app can autonomously decide what context is missing and pitch plot points in real time.

Memory, routing, and proactive chat

Alongside tool selection, the update adds three critical mechanics to Opal workflows.

First is cross-session memory. Opals can now retain user preferences, brand identity guidelines, and ongoing tasks across sessions. TechCrunch noted that the agent can even log data into Google Sheets to maintain long-term state, like a persistent shopping cart.

Second is dynamic routing. Builders can define logic where the agent chooses the next step on the fly. In an executive briefing tool, the agent checks if a client is new or existing, deciding whether to scour the web for background research or dig through internal meeting notes.

Third is interactive chat. If a prompt lacks critical details, the agent pauses execution and actively asks clarifying questions or presents options before moving ahead.

Opal first debuted for US users in July 2025 as a vibe-coding playground for mini web apps. It expanded to over 160 countries by late 2025 and arrived inside the Gemini web app in December.

Why it matters

For non-technical creators, this changes the game. You no longer need to sketch complex conditional diagrams just to build a responsive app. The ability to give an app long-term memory and interactive follow-ups lowers the barrier to creating bespoke, practical software.

Across the AI landscape, Google is positioning Opal directly against red-hot vibe-coding and agent platforms like Lovable, Replit, Emergent, and Rocket.new. Google's core advantage is native plumbing: it seamlessly hooks Gemini 3 Flash into Search, Veo, and Google Sheets without external API keys or configuration headaches.

We think Google made a clever architectural compromise here. Rather than turning the entire workflow into an unpredictable black box, it keeps fixed steps for granular control while letting users isolate autonomous behavior within specific agent nodes.

The potential pitfall, however, lies in agent hallucination during dynamic routing. When an agent autonomously decides which branch to trigger, subtle misinterpretations could lead to circular question loops or unintended tool executions that frustrate users.