Holo3: New Record in AI Agents Operating Computers

While most AI news focuses on chatbots and text generation, another race is quietly unfolding: the race for AI agents that can do more than just answer questions; they can actually operate a computer. This involves opening applications, clicking buttons, filling out forms, and completing tasks within a real interface – just as a human would.

Company H has introduced Holo3, a new version of its model specifically focused on this task. And based on the test results, this isn't just another release for the sake of it.

What is OSWorld for AI Agents and Why Does It Matter?

What is OSWorld and Why Does It Matter?

To understand why Holo3's result is so interesting, we first need to look at how an AI's ability to operate a computer is measured.

There's a special benchmark called OSWorld-Verified. In short, it's a set of real-world tasks on a standard desktop: opening a file, finding specific information, performing an action in a browser or an office application. The model has to «see» the screen and perform the steps on its own – without prompts, without special adapters, and under conditions as close as possible to real-world work.

This is fundamentally more complex than solving math problems or writing code in an isolated environment. There's no clear «right answer» – instead, there's a real interface that can behave unpredictably, and a task that needs to be seen through to completion.

Holo3 OSWorld Score Explained: Is 78.85% Significant?

78.85% – Is That a Lot or a Little?

Holo3-122B-A10B scored 78.85% on OSWorld-Verified, setting a new record among all known models on this benchmark.

For comparison, just a few months ago, the results of the best agents on similar tasks were much more modest. The race here is moving at a breakneck pace – much like the language model race accelerated last year once it became clear that no single company holds a monopoly on progress.

The 78.85% figure means the model can handle almost four out of five tasks in a real desktop environment. The remaining ~21% is where things go wrong: a non-standard interface, an unexpected sequence of actions, or an edge case.

Simply put, it's no longer a «demonstration toy», but it's not yet a tool you can trust with anything unsupervised.

Autonomous Enterprise Concept: How AI Agents Transform Businesses

«Autonomous Enterprise» – What's the Idea Behind It?

H is positioning Holo3 as part of a concept the company calls the Autonomous Enterprise.

The idea is this: a large part of office work consists of repetitive actions on a computer. Filling out a report, transferring data from one system to another, checking a task's status, responding to a standard request. A person spends hours on this. An AI agent that can operate a standard interface could do it all on its own – without special integrations, without APIs, and without needing a separate script for every task.

This is fundamentally different from the «connect the AI to your database via an API» approach. The agent simply looks at the screen and acts – like a new employee who has just been shown their work computer.

AI Computer Agents: Current Trends and Future Landscape

Why Now, and What's Happening in This Niche?

Interest in computer agents has surged in recent months. This is no accident.

First, language models have reached a level where they can reliably understand instructions and context, allowing them to be trusted with multi-step tasks. Second, methods have emerged that enable the model to «see» the screen and interpret the visual interface, not just the text.

OpenAI is moving in the same direction: GPT-5.4, released in early March 2026, was introduced as the company's first model with built-in support for operating on a user's computer in agent mode. Alibaba discovered that in its multimodal Qwen3.5-Omni, the model could write code by watching a screen recording – and this ability wasn't intentionally built-in; it emerged on its own.

In other words, several major players are converging on the same point, but from different angles. H is taking a direct approach – with a specialized model trained specifically for desktop control.

Practical Implications of Advanced AI Agents for Computer Operation

What Does This Mean in Practice?

In short, not much yet, but the trajectory is clear.

A 78.85% score on a benchmark is not the same as «working in a real company.» The test is carefully structured, and the conditions are reproducible. A real office is different: legacy software versions, non-standard configurations, and tasks that are never explicitly formulated.

But results like these indicate that the technology has moved from the «interesting experiment» stage to the «this can already be used in controlled environments» stage. The next step is to expand these environments to something more closely resembling a real-world workspace.

For developers and companies monitoring business process automation, this is a signal: agents that can operate a standard computer interface are no longer science fiction or a distant future. This is a rapidly developing niche where results change literally every few weeks.

The question of reliability remains open: how well does such an agent cope when things don't go according to plan? How does it react to an error? Can it stop and report a problem instead of continuing down the wrong path? These are the things that are still difficult to measure with a single number – and they will determine the real-world applicability of such systems.