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

AI agents are currently one of the hottest topics of technology – but how many enterprises are really posted and actively used?

Linkedin He says he is with him LinkedIn Hiring Assistant. Popular recommendants and AI’s employed search employs business candidates through the company’s AI agent sources and a simple natural language interface.

“This is not a demo product,” said Deepak Agarwal, the main AI officer in Linkedin, said that it was around this week Turn a vb. “It’s alive. It takes a lot of time for employers so that they can take time to do their time to do their time and do their time to do their favorite to hire the best talent for work.”

>> – siSee all transforms 2025 coverage of our entire transform<

Relying on a multi-agent system

Linkedin It takes a multi-agent approach using the things that are described as a collection of agents who collaborate with Agarwal’s work. The supervisor agent combines all the tasks among all the tasks between other agents, including “one and only one and only one job”.

All communication roles occur through the controller agent, which includes human users around specialties and other details. This agent then provides the context of a welding agent, which is encountered with drawings that can fit well together with the employer agent, employer search stacks and sources candidates. After that, this information is returned to the supervisor agent who started interacting with the human user.

“Then you can cooperate with him, isn’t it?” said agarwal. “You can change it. You don’t have to talk to the keywords platform. You can speak natural language with the platform and answer you, it intends to talk to you.”

The agent can then clean the specialties and start candidates for candidates working for both synchronous, synchronous and asynchronous “employees. “He knows what time you will entrust the task agent“How to collect the feedback and show the user,” Agarwal said.

He always stressed the importance of “man’s first” agents that always controlled. The goal is to “personalize deep personalization” experiences with AI, who adapt to choices, and continue to develop and develop users’ interaction with it.

“It is to help you carry your work better and more effectively,” Agarwal said.

How to teach LinkedIn’s multi-agent system

Multiple agent systems require an approach to the NUSTANS. LinkedIn team consumes a lot of time to make each low agency effective for a special task to improve the reliability, delicate adjustment and a low loading agent, Tejas Dharamsi, Linkedi’s large workers’ program engineer.

“We make the domain adapted models thin and make them smaller, smart and better,” he said.

Whereas Supervisor Agent It is a special agent that should be highly wise and adaptable. LinkedIn’s bonfire agent can cause the company to use large language models (LLMS). In addition, reinforcement combines learning and sustainable user feedback.

In addition, the agent has an “experienced memory” agarwal explained, so it can store information from the last dialogue. The user can protect long-term memory on choices, and in this process can be discussed that may be important for retreating later.

“Along with experienced memory, global context and intelligent routing, the controller agent is the heart of the agent and continues to improve by learning strengthening.”

Agent iterating throughout the development period

Dharamsi, with AI agents, stressed that the delay point. Before placing your production, LinkedIn model builders must understand how much you can support how many qualifiers in seconds (QPS) models can support and strengthen their blessings. To determine these and other factors, the company gives a lot of results and is the appraisal along with a ntening red team and risk assessment.

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

After placing a Ui prospect, Dharamsi described Linkedin as the AI ​​Agent Platform as “LEGO blocks as the AI ​​developer” as the AI ​​developer could join. ”

“Try the development of agents in Linkedin, how we standardize the development of agents in LinkedIn, so you can repeat them in a row,” he said. Engineers can instead, more information, optimization and damage and prize function than the main recipe or infrastructure.

LinkedIn, in RL, the various algorithms in control, to use the box, management, pruning, pruning, pruning, without concern, began to start the algorithms and trainings.

When building its models, LinkedIn, reliability, confidence, privacy, personalization and price focus on several factors. Models must provide consistent performances from exiting. Users want to know they can trust agents to be consistent; that their work is reliable; used for personalization of past interactions; And these costs do not skyrocket.

“We want to do things that bring happiness to the user to give more value, better perform and hiring their work,” Dharamsi said. “Employers want to pay attention to the source of the correct candidate without spending time in search.”