At Narada AI, our mission is clear: to build intelligent agents that can reliably automate enterprise workflows, securely, autonomously, and at scale. These workflows often span both API and web interfaces, and achieving robust automation requires high precision across both modalities.

We're excited to announce that the Narada Operator has achieved state-of-the-art results on the WebArena benchmark and WebVoyager benchmarks, recording a 64.16% and 97.45% task success rate across benchmark domains, the highest among all autonomous web agents evaluated to date, outperforming major competitors including IBM CUGA and OpenAI Operator.

This milestone is more than a research win; it's a demonstration that enterprise-grade agentic automation is no longer a future vision. It's here.

What is WebArena? Why does it matter?

WebArena is a large-scale benchmark designed to evaluate how well autonomous agents perform in realistic, dynamic web environments. It includes over 800 long-horizon tasks that reflect real-world workflows such as:

• Managing e-commerce storefronts

• Moderating online forums

• Administering internal project management systems

• Handling CMS-based content operations

• Navigating across multiple websites in sequence

Unlike static benchmarks or single-site flows, WebArena demands multi-step planning, grounded reasoning, and flexible behavior. Agents must navigate GUIs through natural language instructions, not APIs, and complete tasks that often require recovery from ambiguity or error.

In its initial release, even GPT-4-based agents achieved just 14% success, underscoring the challenge. The Narada Operator achieved 64.16%, significantly outperforming existing Compute-Use Agents (CUAs) such as IBM CUGA and OpenAI Operator.

Below is a breakdown of Narada Operator's performance WebArena:

What is WebVoyager? Why is it important?

WebVoyager is a large web benchmark including over 600 tasks over 15 different websites, testing an agent's ability to adapt to drastically different real-world environments with tasks including:

• Computing mathematical results from WolframAlpha

• Performing operations on Github

• Navigating to a destination via Google Maps

Since WebVoyager is executed on live sites, it also tests an agent's ability to deal with interruptions scattered throughout the internet, from advertisements to popups. This tests reliability in a dynamic environment.

Narada's Operator sets a new record on WebVoyager, achieving a state of the art result of 97.45%, far above OpenAI Operator's 87% and Browser Use's 89.1%.

How We Achieved It: Agentic Process Automation for Enterprise

Narada's R&D is led by world-class researchers with a track record of impactful AI work, including LLM Compiler (ICML 2024) and Plan-and-Act (ICML 2025). These innovations form the foundation for Narada's unique ability to execute long-horizon tasks with reliability and precision.

A key breakthrough in our architecture is a mechanism for real-time error detection and correction, inspired by how ECC (Error Correction Code) modules ensure computational reliability in CPUs and GPUs. Even when small execution errors occur, Narada Operator is designed to detect, recover, and proceed, ensuring end-to-end task completion.

We also developed a custom planning system that compiles user intent into actionable workflows, integrating error correction directly into the execution plan.

This approach proves especially powerful in multi-site tasks, WebArena's most complex category, where Narada significantly outperformed other state-of-the-art agents like IBM CUGA.

Built for Secure Enterprise Deployment

Narada Operator is a production-grade agent deployed in sensitive enterprise environments where security, privacy, and reliability are essential.

We've designed our system with features tailored for enterprise automation:

• Zero-trust input handling and takeover modes for sensitive workflows.

• Workflow personalization tuned to each organization's needs.

• Real-time fallback strategies for ambiguous or dynamic conditions.

• Full monitoring and replayability for every execution session.

Unlike general-purpose systems, Narada doesn't aim to serve every use case. Our focus is narrow but deep on enterprise Computer Use tasks. Our operator is fine-tuned for specific enterprise workflows and delivers high accuracy, even across long, multi-application environments.

Crucially, Narada delivers this result with enterprise-grade security. We never train on user data and are fully HIPAA, GDPR, and CCPA compliant, and SOC 2 Type II certified. Our agent is currently being used by hyperscalars in the finance, healthcare, and banking sector.

What's Next

While these results are a major milestone, they are just the beginning. We are actively expanding the Narada Operator's capabilities in areas that matter most to enterprises:

• Task complexity: Longer workflows and deeper nested operations.

• Execution speed: Fast, parallel-safe automation.

• Autonomous recovery: Self-correction and retry strategies.

• User preference modeling: Behavior adaptation based on organizational norms.

Our vision is to make software agents trusted members of the enterprise workforce, capable of not just assisting, but executing.

Try the Narada Operator Today

You can try our consumer-grade Narada Operator today via our Chrome extension, just prepend your query with /Operator and let it take over the rest. Please make sure to contact us if you are interested in the Enterprise version of the Operator, which offers more accurate and reliable execution.

Let's redefine what agents can do, together.

November 7, 2025

Share