Public: LinkedIn Hiring Assistant

Architectural notes

Hiring Assistant is LinkedIn's first AI agent (for recruiters), built on
LinkedIn's internal generative-AI/agent platform with a plan-and-execute design and
specialized sub-agents (sourcing, evaluation, outreach).

Named frameworks: LangChain is the primary application framework (chosen over AutoGen,
LlamaIndex, CrewAI) and LangGraph handles complex flows, adapted to LinkedIn messaging
and memory infrastructure.

The orchestration backbone is in-house: an Agent Lifecycle Service, a Skill Registry,
and gRPC/proto3 over LinkedIn's Messaging platform; these have no catalog equivalents.

Models are served behind an OpenAI-compatible API: external GPT models via Azure OpenAI
plus in-house fine-tuned open-source models (Llama) on PyTorch/DeepSpeed/vLLM; exact
GPT versions are not disclosed.

Memory is in-house: conversational and experiential memory plus a Cognitive Memory
Agent (episodic, semantic, procedural layers).

Evaluation and observability: LangSmith for pre-production tracing and evaluation,
OpenTelemetry traces in production; LinkedIn is adopting MCP and experimenting with A2A.

Not publicly documented: a named third-party vector database (retrieval uses in-house
embedding-based retrieval) or exact model versions.