Building Reliable AI Agents Starts Outside the Model

The model is usually the most visible part of an AI agent, but it is rarely the whole system.

The reliability work lives around it: how context is selected, which tools are exposed, how retries are handled, how state is scoped, and how failures are surfaced. A good agent workflow feels intelligent because the surrounding system is deliberate.

For production use cases, I like thinking in terms of narrow responsibilities:

• The retriever should make context auditable.
• The tool layer should be explicit about permissions and side effects.
• The orchestrator should keep state understandable.
• The logs should explain why the agent took a path, not just what it returned.

That is the engineering part I enjoy most right now: taking LLM capability and giving it the shape of a dependable backend system.