Runner H and Surfer-H: A Browser Agent and Its Holo1 Models

H Company has released Runner H, a browser agent that takes a natural-language task and attempts to carry it out on the web. In this June 2025 video, Matthew Berman says he partnered with the company to introduce the launch. He describes the product as free and in beta at the time, and also reviews its research and released model weights.

For a demonstration, Berman asks the agent to find ten Pokémon card listings on eBay and create a Google Sheet with their links. Runner H asks him to connect Google Sheets; he authorizes the integration before the task proceeds. He notes that multiple browser agents can be started in parallel, although the video follows this particular task.

The interface displays a plan, thoughts, notes and web actions alongside a live browser view. Berman watches the agent search for listings and says he will return to the result later. The visible trace makes the steps inspectable, but the interface alone does not establish that every extracted item is accurate.

Berman introduces Surfer-H and the Holo1 family of vision-language models, including navigation and localization demos. Localization means identifying screen coordinates for an element to click. In one example, a model receives a screenshot of a travel site and a hotel-booking request, then suggests selecting an August 6 checkout date for a three-night stay starting August 3.

The [research paper](https://arxiv.org/abs/2506.02865) describes Holo1 as an open-weight model collection and introduces the WebClick evaluation dataset. The weights are available for use and further development, but readers should check the license for the specific model: the [Holo1-3B license](https://huggingface.co/Hcompany/Holo1-3B/blob/main/LICENSE) includes terms inherited from its Qwen base model. Open weights do not by themselves establish that the entire hosted product is open source.

Berman explains the motivation for visual browser agents: many websites are designed for people and do not offer suitable standardized APIs for every task. Rather than build a separate integration for each site, a visual agent can operate graphical controls by clicking, scrolling and typing. This approach can broaden access to web workflows, although it still depends on what the interface exposes.

Surfer-H has three main modules: a policy, a localizer and a validator. According to the paper, it operates from website screenshots without needing the document object model or accessibility tree. A vision-language policy proposes the next action, and the localizer translates a description of a target element into screen coordinates.

Actions such as opening a URL, scrolling, waiting or clicking are executed in sequence. When the policy produces an answer, a validator assesses whether it is suitable. If it rejects the answer, feedback goes into the agent’s memory and execution continues until a result is accepted or a time or cost budget is reached. The modules can use different compatible models.

Berman walks through the paper’s diagram: the task and observations enter memory, the policy chooses an action, and a localizer supplies coordinates when a click or text entry requires them. The browser’s next screenshot feeds the loop, while the validator can return a final answer or send feedback back to memory.

The paper compares Holo1-3B and Holo1-7B with similarly sized models on UI localization benchmarks, including the company’s WebClick dataset. Berman highlights their higher average click accuracy in the displayed table. These are the authors’ results on particular benchmark sets, not a guarantee that either model will correctly click every control on an arbitrary website.

Berman also describes a cost-versus-accuracy chart and cites the company’s 92.2% WebVoyager result at an estimated $0.13 per task, alongside comparisons with other model configurations. These estimates reflect the paper’s inference-cost assumptions and its tested workload, not a universal price for every production task.

The paper’s configuration matters: the 92.2% figure used Holo1-7B for policy and localization but GPT-4o for validation, with up to ten attempts across 643 WebVoyager tasks. It was not an entirely Holo1-based run, despite the wording shown in the video. The paper reports 69.6% on one attempt for that configuration; using Holo1-7B for all three modules produced 80.4% after ten attempts at an estimated $0.06 per task. The standard Holo1 policy also trained on WebVoyager-related tasks, so the score should not be generalized as unseen-task performance.

Returning to the eBay task, Berman opens the generated Google Sheet. It has ten listing rows, but two entries say their links are unavailable because of CAPTCHA. The agent produced a useful partial result after the connection was authorized, but it did not deliver all ten requested links, and the video does not independently verify the remaining listings.

Berman shows options for highly involved, moderately involved or fully automated interaction, along with uploaded documents and a files view. He points to integrations with Google Sheets, Docs, Drive, Notion, Slack and Zapier, which require authentication. He also says payments were marked as coming soon at the time; the video does not demonstrate a completed payment.

Finally, Berman introduces Tester H, then in private beta, as a way to automate website or application quality assurance. The example specifies a natural-language test on an Airbnb page: navigate to a listing, scroll sideways, click the first photo with an orange bed and check that the image appears. He closes by encouraging viewers to explore the released models and tools.