Agentic AI

It helps to treat "agentic" as a dial rather than a label. At the low end sits a deterministic pipeline where a model fills one slot — classify this ticket, extract these fields — and the surrounding code owns every decision. At the high end sits an open loop where the model chooses its own tools, its own order and its own stopping point. Most production systems that work well sit far closer to the low end than the marketing around them suggests, because each degree of freedom you hand to the model is a degree of freedom you have to test.

The engineering question is therefore not "should this be agentic?" but "which specific decision am I delegating, and what happens when it is wrong?" Delegating the choice of which of four documented tools to call is a small, testable delegation. Delegating the decision to spend money, delete data or email a customer is a different category, and usually belongs behind an explicit approval step.

The misconception worth naming is that autonomy and capability are the same axis. They are independent. A highly capable model in a tightly constrained workflow often outperforms the same model in a free-running loop, because constraints remove the failure modes that come from compounding decisions. Autonomy is a cost you pay for flexibility you actually need — not a quality score.