Enterprise leaders have joined a reliable program for nearly two decades. VB transform brings people to develop real enterprise AI strategies. Get more information
AI agents are one of the hottest topics right now in Tech – but how many businesses have really been deployed and are actively using them?
Linked Says that is with her Assistant Having LinkedIn Services. Going beyond its popular proposer system and AI -powered search, the company’s AI agent recruits and recruits job candidates through a simple natural language interface.
LinkedIn his chief AI officer Deepak Agarwal said this week in Steam. VB Transform. “This is live. It is saving a lot of time for recruits so that they just like their time to do what they like to do, who is raising candidates and getting the best skills for this work.”
>> See all our Transform 2025 coverage here <To rely on a multi -agent system
LinkedIn is a multi -agent approach, using Agarwal as a collection of agents that has been supported to perform. A supervisor agent intends all the tasks among other agents, including intake and souring agents who are “good at one thing and are good at one thing.”
All communication is done by the supervisor agent, who inputs human users around the character and other details. The agent then provides context to a sourceing agent, which, along with the piles of recruiting search piles and sources, explaining why they can fit well. This information is then returned to the supervisor agent, which begins actively interacting with the human user.
“Then you can cooperate with it, okay?” Agarwal said. “You can edit it. Now you don’t need to talk to the platform in key terms. You can talk to the platform in a natural language, and it will answer you, it is about to communicate with you.”
The agent can then improve qualification and begin to source candidates, and “work for both harmony and contradictions while working for a hiring manager.” Agarwal said, “It knows when to hand over this task to an agent, how to submit feedback and the user to display.”
He emphasized the importance of “human first” agents that always control consumers. The purpose is to “deeply personal” experiences with AI, which adapts to the priorities, learns from behavior, and is more and more prepared and improves.
Agarwal said, “It’s about helping you meet your work better and more efficiently.
How does Linkedt train their multi -agent system
A multi -agent system requires an important approach to training. LinkedIn’s LinkedIn Senior Staff Software Engineer Tejas Dharmasi explained that the LinkedIn team spends a lot of time making good toning and improving each agent reliably reliable.
He said, “We improve the domain adaptive model and make them smaller, smaller, smaller, and better.”
While the supervisor agent is a special agent that needs to be very intelligent and adaptable. LinkedIn their orchestating agent can argue using the company’s Frontier Language model (LLM). It also includes learning and a permanent user’s opinion.
Moreover, the agent has “experimental memory”, Agarwal explained, so it can maintain information from the recent dialogue. It can also preserve long -term memory about user preferences, and similar discussions that must be remembered later in the process.
“With experimental memory, global context and intelligent rooting, the Supervisor is the heart of the agent, and it is getting better and better by learning,” he said.
Repeating in the Agent Development Cycle
Dharmasi stressed that with AI agents, the delay would have to be kept in important place. Before deploying in production, LinkedIn model builders need to understand how many questions per second (QPS) model can support and how many GPUs need to provide them with electricity. To determine these and other factors, the company operates a lot of estimates and diagnoses it, as well as a red -colored and risk diagnosis.
He said, “We want these models to be faster, and the sub -agents perform their work better, and they are really fast.”
Once deployed, from the UI’s point of view, Dharmasi described the LinkedIn AI agent platform as “Lego Blocks that can plug and play an AI developer”. The abstracts are designed so that users can choose and choose them on the basis of their products and what they want to make.
He explained, “The focus here is how we standardize the development of agents in LinkedIn, so that in a permanent way you can build them again and again, try different assumptions.” Engineers can focus on data, correction and loss and reward function rather than basic prescription or infrastructure.
LinkedIn these engineers provide various algorithms based on RL, monitoring, cutting, quantization and osoon, to use out of the box without worrying about GPU correction or flop, so they can start running algorithms and training.
He said that in developing his model, LinkedIn focuses on a number of factors, including reliability, confidence, privacy, personal nature and price. The models must provide a permanent output without going down. Consumers also want to know that they can rely on agents to remain permanent. That their work is safe. Past communication is being used for personal nature. And its costs are not sky.
“We want to provide the maximum price to the user, do their job better and do things that bring them to happiness, such as hiring services,” said Dharmasi. “The recruiters want to focus on source the right candidate, don’t spend time on the searches.”
