In this talk, we introduce LMOS (Language Model Operating System) - an open-source, cloud-native platform designed to simplify the development and management of multi-agent AI systems. Inspired by traditional operating systems, LMOS abstracts complexities such as memory management, LLM execution, and agent collaboration, leveraging Kubernetes and Istio to orchestrate intelligent agent interactions.

We explore how LMOS addresses key challenges in building multi-agent systems, including scalability, agent discovery and agent collaboration. Through ARC/Kotlin, a DSL for AI agents, developers can rapidly build agents with modular, production-ready architecture. ARC simplifies complex tasks while offering blazing-fast prototyping, seamless tool integration, and high performance through JVM and Kotlin's concurrency constructs.

In LMOS, agents collaborate dynamically based on their capabilities. We'll explore AI-Routing, which powers intelligent decision-making by directing queries to the most relevant agents. LMOS supports two key methods for routing:

• LLM-based approach: Uses language models to analyze and match user queries with the right agent capabilities.
• Vector-based approach: Utilizes vector embeddings to find the most appropriate

One of LMOS's key strengths is its support for channel-based canary rollouts. Using the LMOS Operator, developers can deploy new agent capabilities incrementally with weighted canary releases—gradually introducing updates to a subset of users.