That's because there are dozens of slightly (or significantly) different definitions floating around and everyone who uses the term likes to pretend that their definition is the only one out there and should be obvious to everyone else.
I collect agent definitions. I think the two most important at the moment are Anthropic's and OpenAI's.
An workflow is a collection of steps defined by someone, where the steps can be performed by an LLM call. (i.e. propose a topic -> search -> summarise each link -> gather the summaries -> produce a report)
The "agency" in this example is on the coder that came up with the workflow. It's murky because we used to call these "agents" in the previous gen frameworks.
An agent is a collection of steps defined by the LLM itself, where the steps can be performed by LLM calls (i.e. research topic x for me -> first I need to search (this is the LLM deciding the steps) -> then I need to xxx -> here's the report)
The difference is that sometimes you'll get a report resulting from search, or sometimes the LLM can hallucinate the whole thing without a single "tool call". It's more open ended, but also more chaotic from a programming perspective.
The gist is that the "agency" is now with the LLM driving the "main thread". It decides (based on training data, etc) what tools to use, what steps to take in order to "solve" the prompt it receives.
I think it's interesting that the industry decided that this is the milestone to which the term "agentic" should be attached to, because it requires this kind of explanation even for tech-minded people.
I think for the average consumer, AI will be "agentic" once it can appreciably minimize the amount of interaction needed to negotiate with the real world in areas where the provider of the desired services intentionally require negotiation - getting a refund, cancelling your newspaper subscription, scheduling the cable guy visit, fighting your parking ticket, securing a job interview. That's what an agent does.
It's just a ~~reduce~~ loop, with an API call to an LLM in the middle, and a data-structure to save the conversation messages and append them in next iterations of the loop. If you wanna get fancy, you can add other API calls, or access to your filesystem. Nothing to go crazy about...
Technically it's `scan`, not `reduce`, since every intermediate output is there too. But it's also kind of a trampoline (tail-call re-write for languages that don't support true tail calls), or it will be soon, since these things loose the plot and need to start over.
Giving an LLM access to the command line so it can bash and curl and and python and puppeteer and rm -rf / and send an email to the FBI and whatever it thinks you want it to do.
While it's common that coding agents have a way to execute commands and drive a web browser (usually via MCP) that's not what make it an agent. Agentic workflow just means that LLM has some tools it can ask agent to run, in return this allows LLM/agent to figure out multiple steps to complete a task.
