Because AI is so accessible now, everyone wants to make everything agentic. Throw an LLM at it, write some instructions in a markdown file, and hope it follows them. But LLMs are probabilistic — there's no enforcement. Hope is not a strategy.
The reality is that the best AI workflows live in the middle. Every good flow has a deterministic part and an agentic part. The deterministic part gives you reliability — hardcoded logic, structured data, predictable outcomes. The agentic part gives you flexibility — understanding context, handling ambiguity, making judgment calls.
Not everything needs to be an agent. Most people just need someone to look at their workflow, figure out which parts should be rock-solid code and which parts benefit from AI, and wire it up properly. That's what I do.