Rather than one general-purpose assistant, LinkedIn built a portfolio of narrow, single-purpose internal agents, each trained on a specific slice of internal data:
This narrow-agent architecture depends on curated, task-specific data rather than blanket knowledge-base access — see Golden-Examples Curation Beats Blanket Knowledge-Base Access. Part of the "tools/agents" pillar of Platform / Tools / Culture: A Three-Pillar AI Transformation Framework.
LinkedIn has shipped several narrow agents as concrete instances of this suite:
Each is deliberately single-job rather than general-purpose — see Golden-Examples Curation Beats Blanket Knowledge-Base Access for why LinkedIn avoids giving any one of these broad, unscoped access. A planned Orchestrator Layer for Narrow Agents will let these agents call each other instead of being used only in isolation.
Two data points Cohen cites as evidence the suite has real value beyond pilot theater: (1) LinkedIn's UXR (user research) team started using the growth-prioritization agent to rank which member-facing ideas had the biggest growth opportunity — an adoption pattern the agent wasn't originally designed for, i.e. organic lateral adoption. (2) Retroactively running an old feature spec (the 'Open to Work' green badge) through the newly built trust agent surfaced trust/scam vulnerabilities that weren't caught until well after the feature had shipped — showing the tooling has retrospective analytical value on past decisions, not just forward-looking value on new ones.