The most thoroughly documented multi-agent production system in public. LinkedIn published not just the
product but the whole reusable agent platform beneath it — skill registry, agent mailboxes over
the production messaging stack, memory taxonomy, prompt service with centrally-injected guardrails, and a
billion-scale semantic retrieval layer. Five diagrams here, because four wouldn't hold it.
1.3Bprofiles searched
7named sub-agents
~90%of LLM calls from one agent
81%fewer profiles reviewed
75×EON-8B vs GPT-4 cost
<100msquery embedding p95
Plan-and-execute, deliberately not ReAct
LinkedIn states the rejection reasons
explicitly: instruction-following reliability, hallucinations, and the intelligence-versus-latency tradeoff.
The planner “divides,” the executor “conquers” — and only the executor runs a
ReAct-style loop, inside each plan step.
Sub-agents are tools, not agents
The most-missed detail: “For implementation
efficiency, we model these sub-agents as tools rather than creating separate agent identities and
mailboxes.” Only the top-level Hiring Assistant instance has a mailbox.
Announced Oct 2024, not Oct 2025
A widely-copied database entry gets this
wrong. Hiring Assistant was announced at Talent Connect in October 2024.
Three verbatim constraints worth putting on any
diagram: “the sub-agent never uses customer data across customer boundaries”; memory
“remains scoped to the recruiter's environment and is never used for training LLMs”; and
“actions are audited and reported in the same manner as human users.”
The integration platform is the perception boundary
LinkedIn's own framing: the
unified integrations platform “acts as the perception and action boundary for intelligent
systems, with Hiring Assistant as one of its first major consumers.” Perception = the agent consumes
entities exclusively from the integration core. Action = every write is routed through Integration ID mapping.
L2a
Agent view — supervisor, seven sub-agents, and the gates
Two human-in-the-loop gates sit in the critical path, and one closed feedback loop runs
between evaluation and sourcing. Neither is optional.
This is what makes LinkedIn's case unusual: the Hiring Assistant is one tenant of a general
platform that also runs an SMB variant, a background coding agent, a UI QA agent, an on-call observe agent,
an analytics agent and an incident-investigation agent. Note the two chokepoints — every prompt resolves
through Prompt Source of Truth, and every model call goes through the GenAI proxy. That is where
guardrails live.
“Instead of the calling applications defining
skills over the implementing downstreams, the downstreams define the skill and expose it to the calling
application.” Registration is a build plugin annotating gRPC service definitions with
platform-defined proto3 options.
Agent transport = member messaging
messaging service → DB persistence →
delivery service → agent orchestrator service; nearline retry on failure; nearline retargeting on region
traffic shift. A sync mode was added in Sep 2025 that bypasses the async queue with
“sideways message creation.”
Where models sit on the cost curve
The stated rule, twice: use custom models
only when you have quality, scale or latency constraints. MVP surfaces that scale with #recruiters get
the SOTA hosted model. Evaluation, which scales with #candidates — “orders of magnitude
bigger” — gets fine-tuned small models.
L2c
MUSE — billion-scale semantic retrieval for an agent
Published June 2026. The interesting inversion: all supervision comes from LLM teacher
labels, not engagement signals — and the same prompt-encoded “constitution” governs both
the offline evaluator and the label generator, which is what keeps evaluation and training aligned.
Not a contradiction —
a cohort and date difference (charter cohort Sep 2025 vs all-customer LinkedIn data Jan 2026). LinkedIn never
explains it, so both are shown with their dates. Same for InMail acceptance: 73% / 69% / 66% across sources.
GAPS
What LinkedIn has NOT published
Question
Status
Which LLM backs the supervisor / planner today
Never named. GPT-4o is only named as “during some of this development” for the copilot era.
Identity of the “proprietary LLM” behind Expert Judge and query understanding
Deliberately unnamed. Same for the open-weight reasoning model used as MUSE Teacher.
Which vector store backs experiential / collective memory
Confirmed: “semantic search via vector embeddings.” Store unnamed. Do not draw Venice or Galene here — those are documented only for MUSE profile embeddings.
Are “agent orchestrator service” and “agent lifecycle service” the same thing?
Language is inconsistent across sources. Treated here as one logical orchestration tier.
Has MCP replaced the skill registry in production?
“Incrementally adopting… moving away from” is a direction of travel, not a completed state.
The screening agent's ASR / telephony stack
“Observe, transcribe, and summarize conversations” — no component named.
Cost or QPS of the live agent fleet
Never published.
Sources — all first-party LinkedIn Engineering or recorded conference talks
