LinkedIn built Hiring Assistant, its first production AI agent, for recruiters, and a shared Python framework and agent platform that other LinkedIn teams use to build agents, according to a conference talk by David Tag of LinkedIn's AI team, published on June 13, 2025.
TL;DR: Hiring Assistant takes a recruiter's job description and documents, writes the qualifications, sources candidates in the background, and tells the recruiter when a list is ready. A supervisor agent coordinates sub-agents that call LinkedIn services through "skills". To build it, LinkedIn moved its generative AI work from Java to Python and built a standard service framework. David Tag reports that more than 20 teams use the framework (03:51) and more than 30 services run on it (03:59), and that teams build non-trivial apps in days instead of weeks (09:22). The talk gives no figures for recruiter results.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | YouTube talk (LangChain Interrupt 2025), 15 min |
| Source published | 2025-06-13 |
| Speakers | David Tag, LinkedIn AI team |
| Industry | Professional network and recruiting |
| Function | HR and recruiting, engineering platform |
| Techniques | Supervisor and sub-agents, ambient (background) agent, skill registry, async messaging, layered memory |
| Tools named | Python, gRPC, LangChain, LangGraph, Azure OpenAI, on-premises models |
Independent summary of public material. LinkedIn is not affiliated with Taskade.
The system LinkedIn built
David Tag presents Hiring Assistant as LinkedIn's first production agent. It serves recruiters, and its job is to automate parts of the recruiting process so that recruiters spend more time in conversations with candidates.
The talk walks through one flow:
- The recruiter describes the role, in the demo an experienced growth marketer, and attaches documents about the position.
- The agent writes the qualifications from that input and the documents.
- The agent tells the recruiter that it will work on the search and report back later.
- When the work is done, the recruiter gets a notice and opens a list of the candidates that the agent sourced.
Tag calls this the ambient agent pattern: the agent works in the background and reports back when it has a result. A newer LinkedIn agent for small-business hiring is a separate system, covered in its own entry.
Architecture of the system
Hiring Assistant uses a supervisor and sub-agent design. A supervisor agent coordinates the sub-agents, and each sub-agent calls existing LinkedIn services.
The talk spends most of its time on the platform under the agent:
- Python as the standard. LinkedIn built most business logic in Java. Its first generative AI apps, from late 2022 (04:25), were also in Java, with simple prompts and no conversational memory. Teams wanted Python for prompt work, evals and open-source libraries, so LinkedIn moved generative AI work to Python.
- A service framework. The framework uses Python, gRPC, LangChain and LangGraph, with standard parts for tool calls, model inference on LinkedIn's internal stack, conversational memory and checkpoints. Tag says that a team can switch between Azure OpenAI and on-premises models with a few lines of code (10:01).
- Async messaging. Agents can run for a long time and can depend on each other. LinkedIn extended its existing messaging service to carry agent-to-agent and user-to-agent messages. A queue retries failed messages.
- Layered memory. Agent memory has working, long-term and collective layers. A new interaction fills working memory, and long-term memory grows over more interactions with a user.
- Skills and a skill registry. A skill can be an RPC call, a database query, a prompt or another agent, and an agent can call it synchronously or asynchronously. Teams register skills in a central registry. In the example flow, the supervisor asks the sourcing agent to find a mid-level engineer, the sourcing agent asks the registry for a fitting skill, and then it runs that skill.
- Observability. The team built custom observability for agent runs. Tag says: "You can't fix what you can't observe." (15:03)
Results the source reports
- David Tag reports that more than 20 teams use the service framework, and he calls that a low estimate (03:51).
- He reports that more than 30 services run on it to support generative AI features at LinkedIn (03:59).
- He reports that teams build non-trivial apps in days instead of weeks with community integrations and prebuilt agents (09:22).
- He reports that Java engineers found the agent libraries easy to pick up (09:03).
Critical assessment
The talk gives no result for recruiters, such as time saved, candidate quality or adoption. All numbers are about internal engineering use, and they are the speaker's estimates. The talk took place at a conference run by the vendor of the agent libraries that LinkedIn chose, and the speaker praises those libraries. The source dates from mid-2025 and describes the platform as it stood then. No outside party checked the numbers. The platform design, with a skill registry, async messaging and layered memory, is the most reusable part of the source.
Build this in Taskade
- Build a team of AI agents with persistent memory and multi-agent collaboration that sources and summarizes candidates. Start at AI agents.
- Start from a template in recruiting agents.
- Run the steps around the agents, such as intake and status updates, with HR automations.
- Tell recruiters when a candidate list is ready with the Slack integration, one of 100+ integrations.