In this session, we take a deep dive into
harness engineering and explore how
context and memory engineering techniques combine to build a memory-first agent harness that supports the agent loop, enabling AI agents that learn and adapt from new information.
Agent engineering is producing a constant stream of what appear to be entirely new engineering disciplines. The reality is that clear patterns are emerging across all of them, and understanding those patterns is what separates engineers who build reliable agentic systems from those who are still chasing the latest framework release.
In this session, you will learn the common techniques and patterns from each of these disciplines. We will cover the agent loop in detail, examining how memory reads and writes participate at every stage of the Perceive, Plan, Act, and Reflect cycle. We will explore the three core memory types used in agent systems: episodic memory for tracking what happened, semantic memory for encoding what is known, and procedural memory for capturing how the agent should act.
We will then look at how these memory types map to the five stages of the memory lifecycle: Encoding, Storage, Retrieval, Injection, and Forgetting. From there, we move into harness engineering and sandbox design, before closing with a hands-on coding session where you will implement a memory-first agent harness from scratch.
Attendees can expect to leave with:
- A clear mental model of the emerging engineering disciplines in agent development and how they relate to each other
- A working understanding of the agent loop and how memory participates at every stage
- Practical knowledge of the three memory types and the five-stage memory lifecycle
- Hands-on experience building a memory-first agent harness using Oracle AI Database
- Reusable patterns for harness engineering and sandbox design that can be applied to any agentic system
Duration: 1 hour
