Today, we are introducing Narada Code, a computer-use agent harness that builds, tests, and improves enterprise workflows, alongside OperatorMini, a lightweight execution agent designed to reduce the cost of running them. Together, they help teams turn a task description into an automation that can be refined through repeated execution.

With workflows developed by Narada Code, Operator achieves an 86.1% task success rate on the WebArena benchmark, exceeding the previously reported best result of 74.3% by 11.8percentage points. OperatorMini achieves 83.5%, bringing much of that performance to a lower-cost executor.

How Narada Code Builds Reliable Workflows

Enterprise automation requires more than the ability to click buttons or fill in forms. A workflow must account for how an application behaves, what can go wrong, and how to recover. Building those instructions has often required an engineer to investigate the application, write detailed prompts, and anticipate exceptions. As applications change, that work must be revisited.

Narada Code takes on that development process. It investigates the relevant applications, drafts a workflow, and tests it in a sandbox environment. When a run fails, it inspects the execution, revises the workflow, and tests again. The context gathered during this process helps it develop reusable skills and repair the broader workflow when an interface changes.

For example, consider a recurring task that copies invoice data between disconnected systems. Developing the automation involves learning where to find the required fields, how to enter them in the destination, and how to check that the transfer succeeded. Narada Code applies its investigation and testing cycle to that workflow; Operator or OperatorMini then carries out the resulting instructions.

Narada Code works with existing coding harnesses, including OpenAI Codex and Claude Code, extending them with tools for enterprise automation. By tracing failures, Narada Code can revise the automation approach and heal the automation. Narada Code also incorporates prompting practices and can automatically detect which components from AgentStudio to choose for different parts of the process such as Agentic Selectors, CriticAgent, CoreAgent, and Operator/OperatorMini agents.

This gives teams a way to develop and maintain specialized automations without manually prescribing every execution path. As Narada Code gathers context from testing and execution, it can use that context to improve the workflow and its instructions.

OperatorMini Makes Repeatable Workflows More Economical

OperatorMini needs more explicit step-by-step instructions than the full Operator, as it spends less reasoning tokens. This makes it particularly suitable for pairing with Narada Code since Narada Code is capable enough to provide sufficient and explicit instructions, that can make OperatorMini achieve high accuracy and reliability while being economical.

OperatorMini consumes 6-8x fewer credits than Operator. Its 83.5% WebArena result is 2.6 percentage points below the full Operator configuration and above the previous best reported result. Teams can use OperatorMini for well-defined, repetitive work and retain Operator for tasks that demand more reasoning or involve greater ambiguity.

New Results on WebArena

WebArena tests whether agents can complete tasks in interactive web environments. Its more than 800 tasks cover activities such as shopping, content administration, collaboration, and information gathering, including workflows that span multiple websites. Success depends on completing the requested task through a sequence of actions.

In our earlier Operator announcement, we reported a 64.16% WebArena success rate. With Narada Codedeveloping the workflows, Operator now achieves 86.1%, while OperatorMini achieves 83.5%. These are results for the combined workflow-development and execution configurations.

The reported results exceed the 74.3% score from WebTactix and the 68% score from Claude Code + GBOX MCP, as well as the IBM CUGA and OpenAI Operator CUA results shown below. The 86.1% score is 11.8 percentage points above WebTactix, an approximately 15.9% relative improvement.

The breakdown below covers five websites plus multisite tasks. On multisite tasks, Narada Code with Operator achieves 85.4%, while the OperatorMini configuration achieves 81.3%, compared with 54.2% for WebTactix. These tasks are particularly relevant to enterprise processes that require an agent to carry information and instructions across applications.

Get Started with Narada Code

Narada Code is now available to our enterprise customers, and we are releasing OperatorMini alongside it. Our focus is on helping teams build specialized automations that can be tested, refined, and maintained as their applications evolve.

To explore how Narada Code and OperatorMini could automate a recurring workflow in your organization, book a call with our team.

September 7, 2026

Share