Many brands are adapting to artificial intelligence, the biggest trend of recent years. Web browsers are undergoing one of the most important adaptation processes. AI agents automate the tasks you do with AI and systematically perform pre-set operations on your behalf. The most important areas enabling these are browsers designed for AI agents.
In this written interview where we discussed Opera Neon and AI agents with Opera’s Director of AI R&D, Monika Kurczyńska, we asked the most anticipated questions. Here are the questions and answers:
How Do AI Agents Work?
Is Opera Neon a traditional browser with added AI features, or was it designed as an AI-native product from the ground up?
Opera Neon draws on our 30 years of browser innovation, but it was actually designed from scratch as an AI-agentic browser. That means we didn’t just add AI features to an existing browser. We reimagined the browser’s purpose from the beginning to be action-oriented.
Our architectural strategy consists of a few key elements:
- LLM-agnostic engine: Neon runs on an AI engine we developed in-house and can intelligently select the most appropriate model (like OpenAI, Google) based on the task type.
- Dual execution environment:
- Tasks within the browser (Do) run locally in the user’s own browser session.
- For more intensive generation tasks (Make), Neon creates a virtual machine environment on our servers in Europe and runs the necessary tools there.
- Contextual “Tasks” architecture: Browser tabs are grouped together within task containers. This allows the AI to build a persistent context and memory for the user’s workflow.
- Workflow customization with Cards: Neon also features a tool we call “Cards”. These cards act as tools that grant additional capabilities to prompts. For example, style settings, output format, or specific instructions can be added. Users can create their own workflow cards for frequently performed tasks and explore workflows created by other users via the Cards store.
The idea of AI agents taking action on behalf of the user is a fascinating concept for everyone. How did you design the experience so that users don’t lose control?
We designed Opera Neon not as an uncontrolled bot, but as a partner working alongside the user. Therefore, user control is always maintained.
- The user always chooses the agent: When the browser suggests switching to a more powerful agent (e.g., 1 Minute Research or Neon Make), it always asks for the user’s approval. The user can either accept or continue in chat mode.
- Real-time visibility and control: While the “Do” agent is running, it shows the user the steps it will take and performs the actions live on the page. The user can stop, pause, or completely take back control of the process at any time.
- Contextual awareness: The interface clearly displays which tabs or tasks the AI is referencing. This way, the user can see what information the AI is basing its actions on.
What do you think is the most fundamental change it brings to daily internet usage?
The most fundamental change is the shift from passive browser usage to an active collaboration with AI.
For nearly thirty years, browsers have been passive tools used to access information. Users would switch between tabs, copy and paste, and manually gather and organize information themselves.
Opera Neon changes this approach. The focus is no longer just on finding information, but on directly completing tasks. We call this the “agentic web.” In this model, AI agents understand the user’s intent, manage complex workflows, and not only provide answers but deliver completed results.
When a user tries Opera Neon for the first time, at what moments do they clearly realize it’s not a traditional browser?
This difference usually becomes apparent right from the first use. There are a few reasons for this.
- Start Page Command Center: Instead of a traditional URL bar, the experience begins with the search area on the start page. This area actually functions like a command center. Users can type what they want to do in natural language, upload files, or give voice commands.
- Smart Mode Suggestion: Users don’t have to wonder, “How do I use the AI?” When a request is typed into the search box, Neon automatically analyzes it and suggests the most appropriate agent among Neon Chat, Do, Make, or research agents. This eliminates the confusion of choosing the right tool.
- Tasks Instead of Tabs: Users also encounter an interface different from the classic tab structure. In Neon, navigation is organized around “Tasks”. A task groups relevant tabs together, allowing the AI to analyze and compare multiple sources simultaneously.
- Active Action: The most striking moment is seeing the browser actually take action. Thanks to the “Do” feature, users can watch live within the browser as Neon navigates between pages, fills out forms, or books a trip.
The ability of AI agents to actively perform actions on the web must create challenges in terms of performance, latency, and scalability. What were the biggest technical challenges?
One of the biggest technical challenges was teaching the AI to accurately “see” and interact with dynamic web pages. Moreover, this had to be done without causing latency or creating privacy risks.
For this, Neon analyzes websites through their DOM tree and page layout data. This allows it to understand the entire structure without scrolling the page and interact much faster.
Additionally, for resource-intensive generative tasks—such as building a web app—we offload the processes to our cloud servers in Europe rather than running them on the user’s device. This way, we eliminate hardware limitations.
When AI systems act on behalf of the user, security and data privacy become critical. How does Opera Neon protect user data?
As a European-based company, the approach to privacy is a fundamental issue for us.
- Local execution for sensitive operations: Neon’s “Do” feature runs entirely in the user’s local browser session. Because of this, login credentials, cookies, or payment details are not sent to Opera or third-party AI servers.
- Strict data retention policy: Web page content required for the AI’s action planning is sent to our servers in an encrypted manner and cannot be accessed by anyone. This data is deleted within 24 hours.
- Not used for AI training: We have strict agreements guaranteeing that user data and interactions are not used to train our models or our partners’ models.
- Transparency: Neon clearly displays its planned steps in the side panel, requests user approval for critical operations, and allows users to intervene at any time. Furthermore, Neon’s web page analysis feature can be completely disabled by the user.
{{user}} {{datetime}}
{{text}}