Stop Calling Everything an Agent
The word has lost all meaning — and that should worry anyone building real AI systems.
Every AI startup pitch deck in 2026 contains the word "agent." It is on slide three, slide seven, and somehow also in the team bios. But what does it actually mean anymore?
From a precise term to a vibe
A few years ago, "agent" meant something specific: a system that takes actions in an environment, observes the result, and adjusts. Now it means anything from "an LLM with a button" to "a system prompt with delusions of grandeur."
I am not arguing for semantic purity for its own sake. I am arguing that the imprecision is starting to cost us — engineering teams build the wrong thing, customers buy the wrong promise, and investors fund the wrong moat.
A simple test
Before you call something an agent, ask:
- Does it make decisions you cannot enumerate in advance?
- Does it have a loop, not just a chain?
- Could it surprise you in a useful way?
If the answer to all three is no, you have a workflow. That is fine. Workflows are valuable. They are also easier to build, debug, and trust. Stop selling them as something they're not.
The best agents in production today are 90% workflow with a small, well-defended core of actual autonomy.