The world of generative artificial intelligence is all about them: agents. And for good reason: this type of AI, capable of carrying out a series of relatively complex tasks autonomously, must open up a whole new range of use cases for a technology that still lacks them today.
In the race for this so-called “agentic” AI, the French start-up H placed its pawns in May, thanks to a record fundraising of more than 200 million euros. This Tuesday, it delivers its first agent, called Runner H, capable of browsing the web and the different software of a company to carry out a wide range of tasks. All this, independently, under the simple supervision of a human.
Until now, sector leaders OpenAI, Anthropic and others Google have focused on creating very large language models, the famous LLMs. The agents go a step further: it is the orchestration – by other AI architectures – of these LLMs so that they have a form of decision-making power. The result is AI models with better reasoning, collaboration and planning capabilities.
From there to seeing the future of technology? With the launch of Runner, H claims to have taken another step towards “ artificial superintelligence », this AI capable of a level of reasoning at the human level, dreamed of by the leaders of the sector. But for now, it will focus on an almost equally futuristic objective: providing companies with a “ virtual agent team ».
OpenAI’s new “o1” AI, a first step towards superintelligence
Unlocking the Utility of Generative AI
« The fact that the hype [l’attention, NDLR] pivoting from LLMs to agents is a good thing, because we are finally talking about concrete uses in business. Agents can be the trigger for transforming the LLM bubble into a more sustainable phenomenon », Enthuses Jean-Baptiste Bouzige, CEO of Ekimetrics, a company specializing in the deployment of AI on an industrial scale.
« Many players, like us, are starting to create agents because it is the most obvious way to derive value from LLMs. Instead of giving companies a tool, we give them a result based on the objectives they have set », he adds.
His speech joins that of H, who thus promises “ automate complex and repetitive task flows » thanks to its agents. But the start-up goes further by talking about the inauguration “ of a new era of AI, one that comes into action in the real world ».
The first concrete example given by the start-up concerns recruitment. The agent can post job offers, compile CVs and send personalized reminders to candidates. Better still, it can enrich applications with information found online (for example on LinkedIn). In short: the young shoot promises to “ reduce weeks of work to moments » and this throughout the recruitment process, from the opening of the position to the integration of the new arrival. In theory, the recruiter finds himself relieved of all administrative work and only has to concentrate on the interviews and his decision-making. In other words, on the human part.
This example is only the first in a long list, the young company promising to make many heavy business processes extremely fluid, such as quality assurance of e-commerce websites – which consists of checking their proper functioning and guarantee their maintenance – or even the management of billing and reimbursement flows in healthcare, today often entrusted to external companies.
Artificial intelligence takes action
Promises to keep
« The objective of Runner H is to allow everyone to free up time for creativity, innovation and human relations », summarizes Charles Kantor, co-founder and CEO of H.
These promises are not new, and they even directly echo the discourse already carried by the wave of AI assistants like Microsoft's Copilot. A first wave with relatively mixed success for the moment, the fault of the difficult calculation of the balance between productivity gains and additional costs.
But there is a major difference, at least theoretical, between the two tools. Unlike the assistant, the agent can take charge of a series of actions from a simple simple instruction, where it would be necessary to guide the assistant step by step, and detail the procedure to follow at the same time. written or orally. Only the confrontation of agents with concrete business cases will reveal their effectiveness or not.
To fulfill its promises, H has developed its own small LLMs, with superior performance to equivalent models on the market. In other words, the start-up controls all the pieces of the puzzle, where many agent creators will remain dependent on other companies for the core of the tool.
“Using AI is a matter of life and death for a company” (Emilie Sidiqian, Salesforce)
A new star in French AI
H will need all these assets to establish itself in a field that promises to be very competitive, where technological giants like Salesforce are already developing their own integrated agents. “We are only at the beginning. The possibilities are endless and we intend to go beyond the scaling laws we know today,” ambitions the leader in his press release.
With this coup, H could place itself in the spotlight of the global AI scene. A status that only Mistral has managed to achieve in France. But the start-up gave some scares. In August, it suffered the departure of three of its five founders, Karl Tuyls, Daan Wierstra and Julien Perolat, all known for their work at Google DeepMind, particularly on agents. The French duo composed of CEO Charles Kantor (former researcher at Stanford University) and his Director of New Technologies Laurent Sifre (also ex-Google Deepmind) found himself alone at the helm of a team of more than 40 engineers and researchers.
The start-up announced “ operational differences » on LinkedIn, and the matter ended there. The launch of Runner H barely 6 months after the creation of the company validates the ability of managers to stay the course. And with its XXL financing, it has the means to achieve its ambitions, on a global scale.
