This session explores the evolution from simple LLM chains to robust cognitive architectures, focusing on deploying stateful agents within your own perimeter (On-Premise). We will compare modern orchestration frameworks, contrasting Google ADK (Agent Development Kit) for structured, model-agnostic agent design against LangGraph for granular state control and CrewAI for multi-agent role-playing. A key focus will be the Model Context Protocol (MCP) — the new standard for connecting agents to internal data and tools without vendor lock-in. We will demonstrate how to build a flexible runtime using Vercel AI SDK Core (deployed in Docker) to serve these agents, ensuring full independence from cloud providers. Finally, we will cover the AgentOps stack, detailing how to implement self-hosted observability (via Langfuse) and access policies to safely manage autonomous systems in production.

Many developers create a separate layer for working with APIs, but Nikita believes that this approach is inefficient — it is better to generate code from your API contract. The talk will cover the following topics: how does SDK differ from API, and why is SDK preferable for development? How can you automatically generate an SDK for Swagger and GraphQL? What should you do if you don’t have an API contract? And finally, why is it important for backend developers to understand React? The talk will help optimize development processes and simplify team interactions, making working with APIs more predictable and convenient.