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

AI agents are one of the hottest subjects of technology at the moment – but how many companies have really deployed and actively use them?

Liendin said he has with his Linkedin job assistant. By going beyond its popular recommendation systems and its research fueled by AI, AI agent sources of the company and recruit candidates for employment thanks to a simple interface in natural language.

“This is not a demonstration product,” said Deepak Agarwal, LinkedIn AI head VB transform. “It’s live. It saves a lot of time for recruiters so that they can spend their time doing what they really like to do, which feeds candidates and hiring the best talents for work. ”

>>See all of our Transform 2025 coverage here<

Relying on a multi-agent system

Liendin Adopt a multi-agent approach, using what Garwal has described as a collection of collaborating agents to do the work. A supervisor orchestrates all tasks among the other agents, including admission and supply agents who are “good in one and a single work”.

All communication occurs by the supervisor agent, who takes comments from human users around role qualifications and other details. This agent then provides a context to a supply agent, who leads to research batteries and recruiters’ research sources as well as descriptions on the reasons why they could be a good adjustment for work. This information is then returned to the supervisor’s agent, who begins to actively interact with the human user.

“Then you can collaborate with him, right?” said Agarwal. “You can modify it. You no longer have to talk to the platform in keywords. You can talk to the platform in natural language, and that will answer you, it will have a conversation with you.”

The agent can then refine the qualifications and start to obtain candidates, by working for the recruitment manager “both in a synchronous and asynchronous manner”. “He knows when to delegate the task to what agentHow to collect comments and display the user, ”said Agarwal.

He highlighted the importance of “human” agents who still maintain users. The objective is to “deeply personalize” the experiences with AI that adapt to preferences, learn behavior and continue to evolve and improve more than users interact with it.

“It’s about helping you do your work in a better and more effective way,” said Agarwal.

How LinkedIn trains its multi-agent system

A multi-agent system requires a nuanced approach to training. The LinkedIn team spends a lot of time refining and making each agent effective downstream for its specific task to improve reliability, said Tejas Dharamsi, LinkedIn software engineer.

“We refine the models adapted to the field and make them smaller, smarter and better for our task,” he said.

While the supervisor is a special agent who must be very intelligent and adaptable. The agent orchestrating of LinkedIn can reason by using the large language models (LLM) of the company’s frontier. It also incorporates learning to strengthen and continuous user comments.

In addition, the agent has an “experiential memory,” said Agarwal, so that he can keep information from the recent dialog box. It can also preserve long -term memory on user preferences, and discussions that could be important to recall later in the process.

“Experienced memory, as well as the global context and intelligent routing, is the heart of the supervisor, and it continues to improve thanks to the learning of reinforcement,” he said.

Iterer throughout the agent development cycle

Dharamsi stressed that with AI agents, latency must be on the point. Before deploying in production, LinkedIn model manufacturers must understand how many queries per second (QPS) can take care of and how many GPUs are necessary to feed them. To determine this and other factors, the company manages a lot of inference and makes evaluations, as well as the assessment of reds and risk assessment.

“We want these models to be faster, and the sub-agents do their tasks better, and they are very fast to do so,” he said.

Once deployed, from the point of view of the user interface, Dharamsi described the AI ​​agent platform of Linkedin as “Lego blocks that a developer AI can connect and play”. Abstractions are designed so that users can choose according to their product and what they want to build.

“The emphasis here is how to standardize the development of agents at LinkedIn, so that, in a coherent way, you can build them again and again, try different hypotheses,” he explained. Engineers can rather focus on data, optimization, loss and reward function, rather than the underlying recipe or infrastructure.

Linkedin provides different RL -based engineers, supervised fine adjustment, pruning, quantification and distillation to be used out of the box without worrying about the GPU optimization or flops, so that they can start performing algorithms and training, Dharamsi said.

During the construction of its models, Linkedin focuses on several factors, including reliability, confidence, confidentiality, personalization and the price, he said. The models must provide coherent outings without derailing. Users also want to know that they can count on agents to be consistent; that their work is secure; that past interactions are used to personalize; And these costs do not skyrocket.

“We want to bring more value to the user, to better do their job and do things that bring them happiness, such as hiring,” said Dharamsi. “Recruiters want to focus on supplying the right candidate, not spending time in research.”