In this session, we’ll walk step-by-step through the process of building AI agents in Kotlin using Koog — a lightweight framework for constructing tool-using LLM agents. Starting from first principles, we’ll examine how an agent interacts with a language model, how tools are registered and invoked, and how to structure reliable interactions whether you're working with local models or remote APIs from OpenAI and Google.

Along the way, we’ll explore integrations with Model Control Protocol (MCP) and demonstrate how they enable structured coordination between components. Using Koog's DSL, we'll describe agent workflows as graphs, making the logic of the agent both readable and reproducible.

The session will culminate in building a simple coding agent — capable of generating and testing small programs — showing how to integrate tools like test runners, linters, and file systems into a coherent agent behavior.

Koog is the result of our internal efforts at JetBrains to build a practical, Kotlin-native agent framework, and is already at the core of the coding agent inside the AI Assistant plugin. This talk will be useful for engineers interested in understanding and building LLM-based automation, from scratch to working prototypes, with full control over the process.