I treat applied AI as an extension of engineering, not a separate category with its own rules. An agent in production still needs error handling, clear autonomy limits, observability, and a plan for when it's wrong - exactly like any other system component.

In the Salesforce context, that means judging carefully where Agentforce and AI automations solve a real problem (triage, summarization, first response) and where they add a layer of unpredictability the business never asked for. Not every automation needs an agent; sometimes a deterministic Flow is still the right answer.

The AI Associate certification marks the formal start of this path, but most of the practical learning comes from implementing and watching what breaks in production - the same reasoning I apply to any other technical area.

It's also what I write about most on Force Tricks: Agentforce, Einstein, and agent engineering, always filtered through "did this actually work" rather than theory.