ai-agents
The Prompt Is Dead, Long Live the Prompt
The terminology turns over every six months. Underneath every agent sits the same unspectacular thing: a well-written prompt that decides for itself which tool to reach for next.
People declared the prompt dead, and then quietly gave it new names.
Agent, skill, workflow, tool use — the terminology turns over every six months, and by now each term has a dozen definitions that politely contradict one another. Underneath, every time, sits the same thing: a prompt that decides for itself which tool to reach for next and where it fetches its inputs. A web search, a calculation, a database lookup — the prompt picks whatever it needs. That is the whole difference between an agent and a chat window.
The rest is the question of how such a prompt gets going in the first place, and there are no more than a handful of ways. The diagram above shows six of them. Someone types something. Another system calls it through an interface. A schedule starts it at three in the morning, with nobody watching. A new file turns up in a folder. A message lands in a queue. Or a person signs off before the next step happens.
What is left is remarkably unspectacular. Most of these agents are built in Python, with libraries that have been around for years, and the actual core is a well-written prompt. The king is dead, long live the king — except that the new king looks suspiciously like the old one.