Stacks/Recruiting

LinkedIn Hiring Assistant

LinkedIn's first agentic product for recruiters is a plan-and-execute supervisor on a message-driven platform, with recruiter-specific agent instances and cognitive memory. Engineering posts describe LangGraph and LangChain adapted to LinkedIn infrastructure, LangSmith in pre-production, OpenTelemetry in production, and a trace-fed evaluation platform. Human-in-the-loop control is treated as a product requirement.

Use case

An AI agent for recruiters that sources, evaluates, and engages candidates across 1.2+ billion LinkedIn profiles, including long-running background pipeline work

Category

Recruiting

Team size

not stated

Orchestrator / framework

Hiring Assistant is a plan-and-execute agent on LinkedIn's message-driven agent lifecycle platform. A supervisor agent plans and delegates; some capabilities are decomposed into sub-agents that are modeled as tools. LinkedIn adapted LangGraph and LangChain to work with its messaging and memory infrastructure via custom-built providers. Invocation can be async (messaging) or a newer sync mode that bypasses the async queue for interactive experiences.

Memory

Individualized cognitive memory tools, including experiential memory and conversation memory, each tied to a recruiter-specific agent instance with its own identity and mailbox. Long-term memories are also curated and refined with big-data offline jobs. Memory stores are siloed, with policy-driven interfaces between them.

Evals

A holistic evaluation platform consumes persisted execution traces for offline evaluations, model regression tests, and prompt-tuning experiments.

Observability

Hybrid: LangSmith for pre-production tracing and evaluation (because many agent components are built on LangGraph and LangChain); OpenTelemetry in production, with structured, privacy-safe spans for LLM calls, tool invocation, and memory usage. Traces feed the evaluation platform.

Monthly cost

not stated

Models

not stated

Observability tools

LangSmithOpenTelemetry

Other tools

GraphQL APIClient-side SDK with server-driven / agent-driven UIMCP (incremental adoption)A2A (incremental adoption)gRPC skill registry (being moved off)

Source

https://www.linkedin.com/blog/engineering/ai/how-we-engineered-linkedins-hiring-assistant

Xiaoyang Gu, Xie Lu, and Daniel Hewlett (Hiring Assistant); Karthik R., Xiaofeng Wang, Tyler Grant, Prasanna Narayanasamy, and Jason Belmonti (GenAI agent platform, 2025-09-10) · Oct 20, 2025