Become a member of the event that is trusted by business leaders for almost two decades. VB Transform brings together the people who build the real Enterprise AI strategy. Leather
AI agents are currently one of the most popular topics in technology – but how many companies have actually used and actively use?
LinkedIn says it’s his LinkedIn Hiring Assistant. Further than his popular recommendation systems and AI-driven search, sources of the company’s AI agent and recruit applications through a simple interface for natural language.
“This is not a demo product,” said Deepak Agarwal, Chief AI officer at LinkedIn, this week on stage on stage VB transformation. “This is live. It saves a lot of time for recruiters, so that they can spend their time doing what they really like to do, what candidates feed and hires the best talent for the job.”
>> View all our transformation 2025 coverage here <Trust a multi-agent system
LinkedIn uses a multi-agent approach, using what AGARWAL described as a collection of agents who work together to do the job. A supervisor agent orchestrates all tasks with other agents, including intake and sourcing agents who are ‘good in one and only one job’.
All communication is done through the supervisor agent, who takes input from human users around rolling qualifications and other details. That agent then offers context to a Sourcing agent, who flows through search stacks and sources from Recruiter, together with descriptions about why they might fit well with the task. That information is then sent back to the supervisor agent, who is actively communicating with the human user.
“Then you can work with it, right?” said agarwal. “You can adjust it. You no longer have to talk to the platform in keywords. You can talk to the platform in the natural language, and it will answer you again, it will have a conversation with you.”
The agent can then refine qualifications and start finding candidates, working for the recruitment manager “both synchronously and asynchronous.” “It knows when the task should delegate to which agent, how to collect feedback and to display it to the user,” said Agarwal.
He emphasized the importance of ‘human first’ agents who always keep users control. The goal is to “personalize” experiences with AI that adapts to preferences, learns from behavior and continues to evolve and improve the more users deal with it.
“It’s about helping you achieve your work in a better and more efficient way,” said Agarwal.
How to train LinkedIn to train his multi-agent system
A system with multiple authorities requires a nuanced approach to training. The LinkedIn team spends a lot of time refining and making every electric agent efficient for his specific task to improve reliability, explained injas Dharamsi, LinkedIn Senior Staff Software Engineer.
“We vote domain-adapted models and make them smaller, smarter and better for our task,” he said.
While the supervisor agent is a special agent who must be very intelligent and adjustable. The orchestra agent of LinkedIn can reason by using the Frontier Large Language Models (LLMS) of the company. It also contains the learning of reinforcement and continuous feedback from users.
Furthermore, the agent has ‘experience memory’, Agarwal explained so that it can store information from the recent dialog box. It can also retain long -term memory about user preferences and discussions that can be important to remember later in the process.
“Experiential memory, along with global context and intelligent routing, is the heart of the supervisor agent, and it gets better and better by learning strengthening,” he said.
Itteren during the agent’s development cycle
Dharamsi emphasized that the latency should be on the point of AI agents. Before they are implemented in production, LinkedIn models builders must understand how many queries per second (QPS) models can support and how many GPUs are needed to feed them. To determine this and other factors, the company has a lot of inference and does evaluations, together with nensive red teaming and risk assessment.
“We want these models to be faster, and sub-agents to do their duties better, and they are really fast to do that,” he said.
Once implemented, from an onion perspective, Dharamsi described the AI agent platform of LinkedIn as “Lego blocks that an AI developer can connect and play.” The abstractions are designed so that users can choose based on their product and what they want to build.
“The focus here is how we standardize the development of agents at LinkedIn, so that you can rebuild them in a consistent way, try different hypotheses,” he explained. Insteads, engineers can concentrate on data, optimization and loss and remuneration function, instead of the underlying recipe or infrastructure.
LinkedIn offers engineers various algorithms based on RL, guided fine tuning, pruning, quantization and distillation to use out of the box without worrying about GPU optimization or flops, so that they can start performing algorithms and training, Dharamsi said.
When building his models, LinkedIn focuses on various factors, including reliability, trust, privacy, personalization and price, he said. Models must offer consistent outputs without getting derailed. Users also want to know that they can rely on agents to be consistent; that their work is safe; that interactions from the past are used to personalize; And those costs don’t shoot up.
“We want to offer more value to the user, to better put their work back and do things that bring them happiness, such as hiring,” said Dharamsi. “Recruiters want to concentrate on purchasing the right candidate, not spending time on searches.”