The Scale-Up presents three AI agents at the forefront of benchmarks in web navigation and publishes Holo-1, an open agental model in the state of art.
We cannot blame him for giving in to fashion effects with agentic AI. And for good reason: H, or H Company, was thought as agentic by nature from its creation. Founded at the end of 2023 by four ex-employees of Google Deepmind and Charles Kantor, a former Centraleupélec and Stanford, the French start-up today announces three new autonomous navigation agents capable of competing with the best American solutions. The Scale-up also publishes in open source on Hugging Face Holo-1, its first Language Model vision, capable of navigating independently.
3 agents, 3 different uses
H therefore presents 3 agents: Runner H, surf H and test H. Everyone responds to very specific use cases. The latter are intended for professionals but also curious individuals to test autonomous agent navigation.
Runner H: an action agent based on connectors
Runner H works as an intelligent orchestration platform which transforms an instruction into natural language into a sequence of coordinated actions on several applications and services. The pilot agent of specialized sub-agents and is based on a system of pre-configured connectors to interact directly with collaborative tools. The platform already allows you to connect to Gmail for email management, Google Workspace (Docs, Calendar, Drive, Sheets) for office automation, concept, Slack for team communication, and zapier to extend the automation possibilities.
H cites several concrete use cases: data synchronization between several platforms, automatic generation of reports from multiple sources, or even complex marketing workflows orchestration involving several tools. The user only has to enter their prompt and the VLM of H is responsible for orchestrating the different actions between the tools.
Surfer H: a web browsing agent
Operator at Openai, Computer User at Anthropic, Mariner Project at Google… Surfer H is positioned in the line of autonomous web agents. Surfer H simulates the behavior of a human user: he “sees” the web page thanks to his vision model (Holo-1), identifies the interactive elements (buttons, forms, menus), and performs the requested actions by clicking, entering text or navigating between the pages.
Surfer H is excellent in autonomous navigation. The agent obtains a Sota score on webvoyer benchmark with a score of 92.2% against 89.1% for Browser Use (direct competitor of H) and 87% for Operator of Openai. Each task costs about 0.13 dollars in material resources, making it one of the most accessible solutions on the market.
Test H: an automation agent for web scripts
Test H allows non-technical teams to create web test campaigns simply by describing the actions to be carried out: “Check that the connection button works”, “Test the order process until payment”, or “Ensure that contact forms send emails well”. Test H then takes care of converting these instructions into a automated web and mobile automated scripts. The agent works like an Automated QA tester: he independently checks the proper functioning of web and mobile interfaces.
Holo1 available in open source, on Hugging Face
In addition to the launch of its three agents, H publishes Holo-1, its engine model (without however revealing the version currently in production), in Open Source on Hugging Face. “Traditionally, an agent uses two separate components: an LLM to plan the steps, go to a site, click, scroller, and a visual model, VLM, to locate the elements. The agents currently on the market are expensive because we must orchestrate these two systems. With Holo 1, we have created a unique model that makes both planning and location, which allows cost reductions Considerable, “explains Charles Kantor, CEO.
The model is available in 3 and 7 billion parameters. The weights of 7B are published under Apache 2.0 license, the most permissive. This latest version reaches the best precision rate on UX navigation tasks among small models. Finally, H also publishes a dataset (webclick) containing 1,639 real interaction scenarios with user interfaces. At the same time, Amazon, present in the capital of H, announces the integration of Holo-1 on the Marketplace AWS. Companies can thus deploy the model in one click on their cloud.
On the availability side, Runner H and Surfer H are already accessible to the public with a progressive deployment. Testing H, it is only available for companies wishing to industrialize their large -scale software tests. With its three agents and the publication of Holo-1, H Company wins slowly but surely in a sector with explosive potential.
