Architecting Reliability in Agentic AI Workflows

Over the last many months, I've been building agentic AI workflows, partly to learn, partly to create something genuinely useful that can be used by others.

There's a lot to share and discuss in this space, and I want to share some early observations about using AI Agents:

. AI Agents are a key part of your workflow. Not the whole thing. It can help you perform tasks, embed logic, absorb large context of data for output, but the architecture around it determines how well it actually performs.

. Quality assurance with AI requires intention and strategy. AI is surface-level capable at QA, but without a well-defined verification plan, you're trusting a non-deterministic system with deterministic expectations. Code follows logic and math. AI doesn't, and tweaking it can produce results you never planned for. How can you ensure what you created is reliable when you have an AI Agent embedded in your code? What's your testing strategy? I find that little to no one is talking in depth about this.

. The structure surrounding your agent matters as much as the agent itself. There will be people who architect clean, reliable frameworks. And there will be people who throw an agent into a workflow, slap some automation on top, and call it agentic AI. Beware of the slop that you'll encounter. If you're not educated in the terminology and concepts to ask the right questions, you're putting yourself in a difficult position when it comes to making decisions on something that will have consequences or downstream effects.

As AI becomes more embedded in critical systems, the people who understand the entire ecosystem, dependencies, models, and pitfalls, will be the ones making decisions that actually hold up. My plan is to keep building, forming best practices, establishing frameworks, reusable flows, and exploring quality assurance to verify code that uses AI elements.