Agentic Architecture in Practice: Harness, LLM, Directives by Rocky Lhotka

## Agentic Architecture in Practice: Harness, LLM, Directives

An agent is not just an LLM with tools bolted on. In practice, it’s the combination of three things: a harness that controls execution, an LLM that reasons and generates, and directives that shape behavior. This talk uses RockBot as a concrete example of that architecture and shows how the pieces fit together in a real system. From there, we’ll look at what becomes possible once the core is solid: memory that preserves context, skills that package repeatable behavior, self-learning and skill evolution that improve the system over time, and integration layers like MCP and A2A that let agents reach tools and collaborate with other agents. The point is to move from abstract “agent” talk to a working model for how agentic systems are actually designed, governed, and extended.

This will be recorded and made available soon after. You can find recordings on https://youtube.com/@tulsadnug