Agentic workflows are where a script is used to provide guidance to the llm on the character it is playing, and therefore the context in which the interaction exists. It’s like a system prompt that changes with the application, like for example you might write a script where one agent is told that it is a phd level research assistant at MIT, The next agent is the head of the department , a critical research reviewer, reviewing the research and pointing out the flaws to the “researcher” personality, that then revises and improves the answer, in a loop, with the reviewer criticising the “research” until finally the reviewer either is satisfied or the “researcher” cannot provide any new insight.

That’s a tiny oversimplification, but basically that. You combine multiple agents with different “expertise” and give them access to a vectorised database of your information, or to tools that can ingest information for llm use, and you have a couple of half drunk interns that you can task with doing stuff for you.

It’s pretty cool, and sometimes also hilarious. I’ve had agents get quite testy with each other and even complain about their compatriots (they are all given names in the script). A little sexual harassment as well. (I often use uncensored models, I find that for critic / skeptic roles they are far superior).

Llms ape the best and the worst of humanity, but fortunately mostly the best parts.