When we build an AI agent, the model is only one part of the stack.

The system’s actual behavior emerges from the code around it: the execution loop, context management, tools, memory, state, retries, permissions, sandboxing, skills, and the mechanisms that decide when to continue, stop, or change strategy. Together, these components form the agent harness.

In this talk, we’ll look under the hood of modern agents, starting from the code and architectural primitives that make them work.

We’ll see how a simple loop evolves into a system capable of executing commands, calling APIs, maintaining memory, managing persistent sessions, acquiring skills, and operating inside isolated environments. We’ll also explore where MCP, A2A, and Google ADK fit into this stack, and what problems they actually solve.

No chatbot “hello world” and no framework API tour. The goal is to build a practical mental model for reading the code of an agent, designing one from scratch, and understanding the trade-offs that emerge when moving from a demo to production.

By the end of the talk, we’ll have a concrete answer to a very developer-oriented question: what actually needs to exist in the code for an LLM to become an agent?