Building an Enterprise-Grade AI Agent for Recruiting at Scale

linkedin · HR · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise recruiting automation

LinkedIn needed a production recruiting agent that could source, evaluate, screen, and learn from recruiter feedback across a massive professional graph while preserving customer boundaries and recruiter control.

Constraints:

The agent had to operate across more than a billion profiles and many asynchronous recruiting workflows.
Recruiter-facing interactions needed real-time conversational behavior while background jobs ran at large scale.
Customer data and learned preferences had to remain isolated by recruiter environment.
Recruiters still needed final authority over hiring decisions.

Architecture

LinkedIn Hiring Assistant is a plan-and-executor agent on a message-driven agent lifecycle platform. Each assistant instance has its own identity and mailbox; a supervisor coordinates planner and executor roles; specialized sub-agents perform intake, sourcing, evaluation, outreach, screening, learning, and cognitive-memory work through headless tools that mirror recruiter workflows.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  recruiter[Recruiter] --> hiring-assistant[Hiring Assistant]
  hiring-assistant --> supervisor-agent[Supervisor Agent]
  supervisor-agent --> planner[Planner]
  planner --> executor[Executor]
  supervisor-agent --> executor[Executor]
  executor --> headless-recruiting-tools[Headless Recruiting Tools]
  headless-recruiting-tools --> profile-graph[Profile Graph]
  executor --> sub-agents[Sub-Agents]
  sub-agents --> cognitive-memory[Cognitive Memory]
  cognitive-memory --> hiring-assistant[Hiring Assistant]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Recruiter`	Human user who supplies goals, reviews insights, and makes final decisions.	`Hiring Assistant`
`Hiring Assistant`	Agent identity and mailbox for a recruiter's workflow.	`Supervisor Agent`, `Recruiter`, `Cognitive Memory`
`Supervisor Agent`	Central coordinator that divides work into plans and execution steps.	`Planner`, `Executor`
`Planner`	Produces structured task-specific plans.	`Executor`
`Executor`	Runs ReAct-style loops and invokes recruiter-like tools.	`Headless Recruiting Tools`, `Sub-Agents`
`Headless Recruiting Tools`	Tools that mirror recruiter actions over profiles, pipelines, and outreach.	`Profile Graph`, `Executor`
`Sub-Agents`	Specialized intake, sourcing, evaluation, outreach, screening, learning, and memory workers.	`Executor`, `Cognitive Memory`
`Profile Graph`	LinkedIn's large-scale candidate and profile corpus.	`Headless Recruiting Tools`
`Cognitive Memory`	Scoped learning layer for recruiter preferences and feedback signals.	`Sub-Agents`

Data flow

A recruiter task enters an assistant mailbox. The supervisor produces a plan, the executor runs ReAct-style tool loops, sourcing and evaluation jobs operate across profile data, and recruiter actions feed a scoped memory system that improves future recommendations without crossing customer boundaries.

“its core, Hiring Assistant is a plan-and-executor agent built on top of a message-driven agent lifecycle platform (see our recent blog for more), supported by”article, stated

“Every click, shortlist, or outreach becomes a signal that sharpens the agent’s understanding of role-specific nuances, preferences, and recruiter style.”article, stated

Why this architecture

LinkedIn separated identity, planning, execution, tools, and memory because recruiting combines real-time dialogue, large asynchronous jobs, individualized preferences, and strict data isolation. Sub-agents as tools avoided proliferating independent mailboxes while still giving the supervisor specialized capabilities.

“real-time conversational interfaces, large-scale asynchronous execution, reliable agentic tool use, and individualized cognitive memory”article, stated

“Some of Hiring Assistant’s capabilities are further decomposed into sub-agents for modularity and abstraction. For implementation efficiency, we model these sub-agents as tools rather than creating separate agent identities and mailboxes.”article, stated

Models used: fine-tuned models · LLMs · ReAct-style executor

Scale

The agent sources and evaluates candidates across more than 1.2 billion profiles.
“sourcing and evaluating candidates across 1.2+ billion profiles with enterprise-grade throughput and reliability.”article, stated
Sourcing and evaluation jobs run across millions of profiles.
“* Massive background execution: Runs sourcing and evaluation jobs across millions of profiles. * Continuous pipeline management: Refreshes candidate pipelines daily and evaluates new applicants continuously.”article, stated
Candidate pipelines refresh daily and new applicants are evaluated continuously.
“* Continuous pipeline management: Refreshes candidate pipelines daily and evaluates new applicants continuously. * Human-in-the-loop: Ensure that humans are in the loop for all decisions that humans are accountable for while removing much of the toil from them.”article, stated
Fine-tuned models evaluate candidates in seconds.
“Using techniques such as speculative decoding within our custom LLM serving infrastructure, our fine-tuned models can evaluate candidates in seconds.”article, stated

Alternatives considered (roads not taken)

Create separate agent identities and mailboxes for every specialized sub-agent.

Rejected because: LinkedIn used sub-agents as tools to keep orchestration under the main assistant identity.

Fully automate hiring decisions.

Rejected because: Recruiters remain responsible for reviewing insights and making final decisions.

“and evaluation rubrics to produce structured recommendations with reasoning and evidence. Recruiters remain in the loop to review these insights and make the final decisions on advancing candidates through the recruiting workflow. * Candidate outreach agent: Handles communication with”article, stated

Share learning across customer boundaries.

Rejected because: Memory and sub-agent behavior must stay scoped to the customer environment.

“Importantly, in keeping with our enterprise grade data isolation and responsible AI principles, the sub-agent never uses customer data across customer boundaries.”article, stated

Patterns

`multi-agent-orchestration` (agent architecture)

A supervisor coordinates planner, executor, and specialized recruiting sub-agents.

“supervisor agent, the “central nervous system””article, stated

`async-message-orchestration` (runtime)

Each assistant instance has an identity and mailbox, and the lifecycle runs through asynchronous messages.

“lifecycle is orchestrated entirely through asynchronous message passing.”article, stated

`agent-memory` (personalization)

Recruiter actions become scoped signals for future assistant behavior.

“All memory data remains scoped to the recruiter’s environment and is never used for training LLMs.”article, stated

`human-in-the-loop-gate` (governance)

Recruiters review agent insights and make final decisions.

“Recruiters remain in the loop to review these insights and make the final decisions”article, stated

Tradeoffs

Sub-agent topology

Chose Sub-agents exposed as tools over Separate sub-agent identities and mailboxes.

Rationale: This kept orchestration centralized while preserving specialized capabilities.

“sub-agents as tools rather than creating separate agent identities and mailboxes.”article, stated

Decision authority

Chose Recruiter-in-the-loop recommendations over Autonomous final hiring decisions.

Rationale: The system assists evaluation and sourcing but leaves final judgment to recruiters.

“Recruiters remain in the loop to review these insights and make the final decisions”article, stated

Learning scope

Chose Customer-scoped cognitive memory over Cross-customer shared learning from recruiter actions.

Rationale: Preference learning had to respect customer data boundaries.

“sub-agent never uses customer data across customer boundaries.”article, stated

Gotchas (what bit them)

Agent memory can become a privacy boundary problem.

Root cause: Recruiter feedback and preferences are customer-specific and sensitive.

Resolution: Scope memory to the recruiter environment and avoid training LLMs on it.

“All memory data remains scoped to the recruiter’s environment and is never used for training LLMs.”article, stated

Large asynchronous jobs must not break conversational UX.

Root cause: Sourcing and evaluation span millions of profiles while recruiters expect interactive workflows.

Resolution: Use message-driven lifecycle orchestration with background jobs and real-time interfaces.

“real-time conversational interfaces, large-scale asynchronous execution”article, stated

Evaluation

LinkedIn evaluates the system through throughput, profile coverage, candidate evaluation latency, daily refresh behavior, recruiter feedback signals, and human review of assistant insights.

Metrics: Profiles covered · Candidate evaluation time · Daily pipeline refresh · Continuous applicant evaluation · Recruiter feedback signals

“Every click, shortlist, or outreach becomes a signal”article, stated

Outcomes

Hiring Assistant supports sourcing and evaluation across 1.2B-plus profiles.
Recruiters stay in the loop while background agents refresh pipelines and evaluate applicants.
Scoped cognitive memory lets the system learn from recruiter behavior without crossing customer boundaries.

The source reports 1.2B-plus profiles, millions of profiles in sourcing/evaluation jobs, daily pipeline refreshes, continuous applicant evaluation, and candidate evaluation in seconds.

“sourcing and evaluating candidates across 1.2+ billion profiles with enterprise-grade throughput and reliability.”article, stated