🧠 From Prompts to Partners: How LinkedIn Built Its Agentic Stack

Inside the tech shift that turned LLMs into collaborative, reasoning systems powering LinkedIn’s new AI agents.

Last year LinkedIn unveiled its generative-AI tech stack, the one that powered products like Hiring Assistant, an AI that helps recruiters shortlist and message candidates.

Now that assistant is going global 🌍 and LinkedIn’s stack has evolved again, this time to support full-blown multi-agent systems.

Behind that jump lies a hard engineering problem:
How do you make dozens of reasoning agents think, plan, and act together while staying safe, observable, and fast enough for millions of users?

---

🚀 Your Chance to Sponsor Byte-Sized Design

We’re opening limited slots for sponsors! If you want your product in front of 39,000+ engineers and tech leaders who actually build and ship things, now’s your chance.
Get in touch by responding to this email or through Passionfroot

---

🧩 What Makes an Agent Different

Traditional software follows rules.
Agents follow intent.

A LinkedIn GenAI agent interprets goals, plans steps, and interacts with APIs and humans. It keeps memory, reasons with LLMs, and knows when to ask for help (“human-in-the-loop”).

Think of it as LinkedIn’s shift from “prompt → reply” to “mission → workflow.”

Each agent is modular—a micro-service with cognition.
That modularity gives LinkedIn four things every distributed-systems engineer loves:

• Scalability ⚡: Run subtasks in parallel.

• Resilience 🧱: Failures isolate cleanly.

• Flexibility 🔌: New skills = new agents.

• Observability 👀: You can trace what it was thinking.

---

🧠 How They Define Agents

Instead of inventing a new interface, LinkedIn reused its production muscle—gRPC.

Developers define agents as standard gRPC services, annotate metadata, and register them in a Skill Registry (a directory of all available agents and how to call them).

When the build runs, a plugin registers that definition automatically.

It’s simple, type-safe, and dev-friendly—a pattern every infra team should envy.

---

🔄 Multi-Agent Orchestration via Messaging

LinkedIn decided to piggy-back on their existing messaging platform instead of reinventing the wheel!