Paris AI lab unveils three execution-focused agents and open-sources visual model

H Company, the Paris-based AI startup founded by former DeepMind scientists, publicly launched three autonomous agents in June 2025 after months of closed beta testing. The move marks the company's shift from research to deployment, with each agent designed to automate specific workflows rather than simply respond to queries.

The three agents reflect a deliberate strategy. Runner H handles orchestration across multiple applications and systems, delegating tasks to specialist sub-agents and coordinating actions across spreadsheets, document systems and web browsers through a single natural language prompt. Surfer H operates as a browser automation agent, navigating web interfaces and executing tasks without requiring custom integrations for each website. Tester H focuses on enterprise software quality assurance, converting written test specifications and user stories into automated test suites for web and mobile applications.

The metric that matters most to customers is cost. Surfer H achieved a 92.2 per cent task completion rate on standard benchmarks whilst costing £0.10 per operation. The company claims this represents a 5-to-10-fold cost reduction versus OpenAI, Anthropic and Google's offerings for equivalent performance. This efficiency gap emerges from H Company's focus on execution rather than conversation. The company argues that chatbot-style AI agents waste computational resources on elaborate reasoning when many workflows simply require reliable action.

"We are witnessing the transition from conversational AI to execution intelligence," Charles Kantor, the company's co-founder, said of the launch. "Entering the era where agents complete entire workflows autonomously, delivering the productivity gains enterprises have been seeking from AI investments."

The performance edge reflects the team's origins. H Company was founded in 2023 by five ex-DeepMind researchers (Laurent Sifre, Karl Tuyls, Daan Wierstra and Julien Perolat) alongside Charles Kantor, a Stanford-trained AI researcher. The company raised £220 million in its seed round in May 2024, backed by investors including Eric Schmidt, Amazon, Accel, Bpifrance and Xavier Niel. The company also appointed Gautier Cloix, a former forward-deployed engineer at Palantir, as chief executive in June 2025 to oversee the shift from research into operations.

The practical advantage of execution-focused agents lies in their deployment simplicity. Traditional enterprise automation often requires custom integrations with each target system through application programming interfaces. Surfer H bypasses this constraint by interacting directly with web interfaces, treating any website as accessible without prior setup. This matters for organisations with complex, legacy software stacks where API access may be limited or expensive to implement.

Holo-1 and open-source strategy

Alongside the agent launches, H Company open-sourced Holo-1, its visual-language model, under an Apache 2.0 licence. The model comes in two sizes: three billion and seven billion parameters. Despite its compact design, Holo-1 achieved 76.2 per cent accuracy on tasks involving interface localization and element identification. It represents the highest performance among open-source models of comparable size.

The release carries strategic weight. By open-sourcing Holo-1, H Company distributes its technology to the broader research and development community whilst positioning itself as the company that built the underlying engine. The company partnered with Hugging Face for model distribution, Nvidia for inference optimization and Amazon Web Services for marketplace availability. It also released WebClick, a dataset of 1,639 real-world user interface interaction scenarios, to help researchers train improved vision-based agents.

For European AI development, the move signals a different competitive strategy than US giants. OpenAI and Anthropic typically retain their most powerful models for commercial product revenue. H Company's choice to release a capable foundation model under an open licence reflects confidence in its execution-focused agents as the actual competitive advantage. The model is valuable, but the agents are the business.

What's next for deployment

Runner H entered public beta immediately, with enterprises able to request access to Surfer H for browser automation tasks. Tester H operates as a private beta product targeted at enterprise software teams, requiring access requests but signalling the company's confidence in demanding use cases like automated testing.

The agents arrive as enterprises gradually abandon the idea that AI primarily means chatbots. The gap between conversational AI and execution intelligence has widened as use cases emerge around workflow automation, data movement between systems and quality assurance. H Company's focus on cost efficiency and compatibility with existing systems places it alongside other specialist players in enterprise automation, rather than directly competing with general-purpose assistants.

The open-sourcing of Holo-1 also leaves unresolved how H Company's commercial model works when competitors can rebuild the underlying model. The company's bet is that execution agents require not just a capable vision model but also training data, orchestration infrastructure and accumulated knowledge about integrating with enterprise systems. Over time, that infrastructure may prove harder to replicate than a model released to the public.