AI agents are currently one of the hottest topics in Tech – but how many enterprises have actually deployed and are actively using them?

Linkedin It is said that it is with it LinkedIn Hiring AssistantGoing beyond their popular recommendation system and AI-operated search, recruit job candidates through the company’s AI agent sources and a simple natural language interface.

LinkedIn Chief AI officer Deepak Aggarwal said, “It is not a demo product.” VB conversion“It’s live. It is saving a lot of time for the recruiters so that they can spend their time what they really like to do, who is nurturing the candidates and hiring the best talent for the job.”

Relying on a multi-agent system

Linkedin Taking a multi-agent approach, Agarwal has used Agrawal described as a collection of agents collaborating to work. A supervisor agent orchens all tasks between other agents, including intake and sourcing agents that “are good in one and only one job.”

All communication supervisors occur through the agent, which takes input from human users input around merit and other details. The agent then provides reference to a sourcing agent, which describes the recruiter search stacks and sources of sources that they give details on why they can be a good fit for the job. The information is then returned to the supervisor agent, which begins to actively interact with the human user.

“So you can cooperate with it, right?” Aggarwal said. “You can modify it. Now you don’t have to talk to the platform in the keyword. You can talk to the stage in natural language, and it’s going to answer you back, it’s going to interact with you.”

The agent can then refine the qualification and start sourcing the candidates, working for the hiring manager “both synchronous and excessively.” “It knows when to hand over the task RepresentativeHow to display the feedback and the user, ”said Aggarwal.

He emphasized the importance of “human first” agents that always keep users under control. The target is “depth personalized” experience with AI that optimize for preferences, learns from behavior and continues to develop and improve the more interacting with users.

Aggarwal said, “This is about helping you complete your job in a better and more efficient way.”

LinkedIn how to train your multi-agent system

A multi-agent system requires a fine approach to training. LinkedIn Senior Staff Software Engineer Tejas Dharmasi said that LinkedIn’s team spends a lot of time on fine-tuning for their specific work and makes each downstream agent efficient for its specific task.

“We fix the domain-friendly model and make them small, smart and better for their work,” he said.

Whereas Supervisor agent There is a special agent that should be highly intelligent and optimal. LinkedIn’s orchestrating agent can argue using the company’s Frontier Large Language Model (LLMS). It also includes reinforcement learning and continuous user response.

In addition, the agent has “experienced memory”, Aggarwal explained, so it can maintain information from recent dialogue. It can preserve long -term memory about user preferences, as well,, and may be important to remember later in this process.

“With experienced memory, global reference and intelligent routing, the supervisor is the heart of the agent, and this reinforcement is better and better through learning,” he said.

Repetition in the entire agent development cycle

Dharamasi insisted that with AI agents, the delay should be at the point. Before deploying in production, LinkedIn model builders need to understand how many queries per second (QPS) models can support and how many GPUs are required to give strength to those. To determine this and other factors, the company guesses and evaluates a lot, as well as with ntensive red teaming and risk evaluation.

“We want these models to do fast, and sub-agent to do their tasks better, and they are really fast to do so,” he said.

Once deployed, from the perspective of a UI, Dharamsi described LinkedIn’s AI agent platform as “Lego blocks that can an AI developer plug and play.” The essence is designed so that the users can choose based on their product and choose what they want to make.

“It is focused here how we standardize the development of agents in LinkedIn, so that in a coherent fashion you can make them again and again, try different hypotheses,” he explained. Engineers can focus on data, adaptation and loss and reward function instead of the underlying recipe or infrastructure instead.

LinkedIn provides engineers with various algorithms based on RL, fine tuning, pruning, perception and distillation are monitored to use out of the box without worrying about GPU optimization or flop, so they can start running algorithms and training, so they said.

In the manufacture of its models, LinkedIn focuses on several factors, including reliability, trust, privacy, privacy and value, he said. The model should provide an output continuously without derailing. Users also want to know that they can rely on agents to be compatible; That their work is safe; That past conversation is being used for privatization; And this cost does not touch the sky.

“We want to provide more value to the user, to do our work back better and do things that bring them happiness, such as hiring,” Dharamsi said. “Recruithers want to focus on sourcing of the right candidate, do not spend time on discoveries.”