LinkedIn has extended its generative AI application platform to support multi-agent systems by repurposing its existing messaging infrastructure as an orchestration layer. This approach allows the company to scale AI agent deployments without building new coordination technology from scratch. The revised architecture enables LinkedIn’s “Hiring Assistant”, the company’s first AI agent, to achieve global availability while supporting complex multi-step workflows through agent coordination.
The engineering team identified several key architectural lessons from their multi-agent implementation:
- Reusing existing infrastructure and providing strong developer abstractions are key to scaling complex AI systems efficiently.
- Designing for human-in-the-loop control ensures trust and safety while enabling agents to operate autonomously when appropriate.
- Observability and context engineering have become essential for debugging, continuous improvement, and delivering adaptive, personalized experiences.
- Finally, adopting open protocols is critical to enabling interoperability and avoiding fragmentation as agent ecosystems grow.
The platform treats agents as standardized gRPC services registered in a central skill registry, allowing developers to define agent capabilities through familiar service contracts. “Developers simply annotate this definition with some platform-defined proto3 options that describe the metadata of their agent, and register it via a build plugin into the skill registry”, creating a reusable agent-as-a-service pattern that leverages existing LinkedIn infrastructure.
The relationship between different agent platform components (source)
Rather than building new orchestration technology, LinkedIn adapted its production messaging system to coordinate multi-agent workflows. This platform offers guaranteed first-in-first-out (FIFO) delivery, message persistence, and horizontal scaling capabilities, and these directly translate to agent coordination benefits. “Long-lived tasks could be broken down into a sequence of messages with guaranteed FIFO delivery and seamless message history lookup”, while parallelization occurs through multiple message threads managed by the existing infrastructure.
The architecture implements a two-tier observability strategy addressing the unique challenges of debugging non-deterministic agent systems. Pre-production environments use LangSmith for rich execution tracing and developer introspection, while production relies on OpenTelemetry for structured, privacy-compliant monitoring integrated with LinkedIn’s existing observability stack. This hybrid approach enables developers to iterate quickly while maintaining enterprise-grade reliability and compliance.
LinkedIn agent platform observability flow (source)
LinkedIn’s client integration libraries handle cross-device synchronization, incremental streaming, and asynchronous communication patterns which are essential for agent interactions that span multiple user sessions. The platform supports both synchronous and asynchronous agent invocation patterns, allowing developers to choose between strong consistency with async messaging or eventual consistency with direct synchronous calls based on specific use case requirements.
The platform adopted open protocols, including Model Context Protocol (MCP) and Agent-to-Agent (A2A). “We are incrementally adopting these open protocols, moving away from a proprietary skill registry, paving the way for more intelligent, interoperable, and context-aware agent ecosystems”, positioning LinkedIn’s architecture for broader ecosystem compatibility as agent standards mature.
MCP has already gained widespread adoption among major model providers, such as Anthropic and OpenAI. However, A2A remains a more experimental protocol, focusing specifically on enabling seamless collaboration between different AI agents. A2A addresses the challenges of agents discovering each other’s capabilities, negotiating task delegation, and coordinating complex workflows that span multiple specialized agents.
This architectural approach reflects a broader industry trend toward enterprise multi-agent platforms, with AWS, Microsoft, and other major providers releasing similar orchestration frameworks in recent months. LinkedIn’s messaging-based approach offers a distinctive path for organizations looking to leverage existing infrastructure investments while scaling AI agent capabilities.
