What "AI agent" means once you strip out the marketing

In the last year, "AI agent" has been used to describe a single prompt-and-response API call, a multi-step workflow with an LLM in one node, and a genuinely autonomous system that plans, acts, and re-plans without a human in the loop. Vendors rarely specify which one they mean.

The distinction that matters for a build isn't philosophical, it's operational: does the system take an irreversible action (sending an email, charging a card, changing a record) without a human checkpoint? If yes, you're building something that needs guardrails most teams underestimate.

Most of what businesses actually need is the middle category: a workflow where an LLM handles judgment-heavy steps (classification, summarization, drafting) inside a deterministic pipeline that a human or a rules engine still controls. That's not a lesser version of an agent, it's the version that ships reliably.

Ask any vendor pitching "autonomous agents" one question: what happens when it's wrong three times in a row? If the answer isn't specific, the system isn't ready for anything that touches money or customers yet.