What enterprise leaders can learn from LinkedIn’s success with AI agents

Artificial intelligence agents is one of the most important topics in technology at the present time – but how many companies have you already published and actively used?

LinkedIn says she has a LinkedIn Employment Assistant. The common recommendation systems and research in which artificial intelligence works goes beyond, the sources of the AI ​​agent in the company and the recruitment of candidates through a simple natural language interface.

“This is not an experimental product,” said Deepak Aguardo, chief artificial intelligence officer in LinkedIn, on the stage at VB Transform. “This is alive. It saves a lot of time for employees so that they can spend their time to do what they really like to do, which sponsors the candidates and employ the best talent for this position.”

>> Watch each coverage of our conversion 2025 here <

Dependence on a multi -agent system

LinkedIn takes a multi -agent approach, using what Agarwal described as a group of agents who cooperate to accomplish the task. The supervisor agent organizes all other tasks among other agents, including input agents and sources “are good in only one job.”

All communications occur through the supervisor agent, who receives inputs from human users regarding the qualifications of roles and other details. This agent then provides a context to a resource agent, which is resorting through the chimneys searching for the recruiter and candidates, as well as descriptions about the reason they are suitable for the job. Then this information is returned to the supervisor’s agent, who begins interacting with the human user.

“Then you can cooperate with it, right?” Agarawal said. “You can modify it. You no longer speak to the basic system in keywords. You can talk to the platform in a natural language, and you will answer you, you will have a conversation with you.”

The agent can then improve qualifications and start obtaining the sources of candidates, and work for the employment director “simultaneously and unspecified.” “He knows when the task must be delegated to any agent, and how to collect comments and offer to the user,” Agarawal said.

He stressed the importance of the “first human” agents that always keep users control. The goal of this is to “deepen” experiences deeply with artificial intelligence that adapts to preferences, learning from behaviors and continues to develop and improve whenever users interact with them.

“It is about helping you in completing your business in a better and more efficient way,” said Aguardus.

How LinkedIn is his multi -agent system

A multi -agent system requires an accurate training approach. Tejas Dharamsi, Tejas Dharamsi, LinkedIn, Senior LinkedIn Summary, LinkedIn, explained that the LinkedIn team spends a lot of time to accurate

“We have models adapting to the field and make them smaller, smarter and better for our mission,” he said.

While the supervisor’s agent is a special agent who requires high intelligence and the ability to adapt. LinkedIn agent can use the company’s large language models (LLMS). It also includes reinforcement learning and continuous user comments.

Moreover, the worker has a “experimental memory”, as AGARWAL explained, so that he can keep information from the last dialogue. It can maintain a long -term memory about user preferences, and discussions that may be important to remember later in this process.

He said: “Experimental memory, in addition to the global context and smart guidance, is the heart of the supervisor’s agent, and it continues to improve and improve through reinforcement learning.”

Repetition during the agent development course

Daramsi emphasized that with artificial intelligence factors, cumin should be at a point. Before spreading in production, the LinkedIn model builds up to understand the number of queries in the second (QPS) and the number of graphics processing units required to operate them. To determine this and other factors, the company runs a lot of reasoning and evaluates, along with the Red Red team and risk assessment.

He said: “We want these models to be faster, and that the Pharaoh agents perform their tasks better, and they are quick to do so.”

Once it is published, from the user interface perspective, DHARAMSI described the AI ​​AI platform from LinkedIn as “LEGO’s blocks that artificial intelligence developer can connect and operate.” The abstracts are designed so that users can choose and choose based on their products and what they want to create.

“The focus here is how we unify the development of agents in LinkedIn, so that you can build these different hypotheses in a repeated way.” Instead engineers can focus on data, improvement, loss and reward, instead of the primary recipe or infrastructure.

Daramsi said that LinkedIn provides engineers with different algorithms that depend on RL, precise control over supervision, trimming, quantities, and distillation for use outside the box without worrying about the improvement of GPU or stirring, so that they can start the algorithms and training.

In building their models, LinkedIn focuses on several factors, including reliability, trust, privacy, customization and price. Models should provide consistent outputs without getting out of their course. Users also want to know that they can rely on agents to be consistent; Their work is safe. Previous reactions are used to customize them; This cost does not rise.

“We want to provide more value to the user, better restore them and do things that bring them happiness, such as employment,” said Darmsey. “The recruits want to focus on the sources of the appropriate candidate, not spending time in searches.”