Today's business post is about the difference between an AI agent and a chatbot, read straight from the two vendor guides that define the terms. Writing it forced a question I had been avoiding. I use the word "agent" for most of what runs my business unattended. This morning I lined each of those things up against the definitions and checked whether any of them is one.
The definition I used
Anthropic's guide puts it in one sentence: workflows are systems where the steps are "predefined code paths", agents are systems where the model "dynamically direct[s] their own processes and tool usage". OpenAI's guide says agents "independently accomplish tasks on your behalf" and that "simple chatbots, single-turn LLMs, or sentiment classifiers" are not agents. So the test is two questions. Does the model choose the steps? And does it decide when it is done?
What is actually running
The marketing engine that drafts these posts. It runs from a numbered list in a file: heartbeat, calendar, catch-up check, sitemaps, post one, post two, images, social, checks, log. It cannot skip a step. It cannot add one. Inside a step it chooses freely, and today it chose to fetch two vendor guides nobody told it to read. It also swapped this post's title, because the calendar's placeholder for today described something I could not prove had happened, and the rule it runs under says a placeholder that the week's work does not support gets swapped and logged rather than written anyway. That is judgement inside a step. The step order and the exit are mine. By the definition, a workflow with an agent inside some of the steps.
The inbox manager. Every two hours it reads my mail, proposes drafts, and archives noise. The task file says "You NEVER send email. You draft replies and archive noise." It never decides that it is finished with a thread; it hands every one back to me. Not an agent. A very good clerk with no signature authority.
The run watchdog. Once a day at half past two it scans a week of heartbeat lines for a day with no entry and reports the gap. The runbook is explicit: "The watchdog alerts; it does not republish." One step, one output. Not an agent.
The video queue. Twenty-two clips scheduled one a day in the social planner, with a schedule file as the record and the daily run confirming today's two rows before it builds anything else. A calendar with a check on it. Not an agent.
The chat widget on my site. It answers from the pages and hands over to a person. By OpenAI's own list, a simple chatbot, and not an agent.
That is the inventory. Nothing on my desk is an agent all the way through. The closest thing is the engine, and the reason it is not one is deliberate, not accidental.
Where the authority actually sits
The business post says the agent versus chatbot difference is authority, not intelligence. My own setup is a demonstration of that sentence. The engine is permitted to read anything public, draft anything, and render anything. Its permission to act stops at a short list of scripted checks that contain no model: a grep for the two dashes I never use, the suburb I never name, and any figure that looks like a price. A hit refuses the payload. The model does not get a vote at that point.
The one time it did take an action that could not be undone was in early August. A test payload without a schedule date published immediately to my Facebook page, and the platform's delete call returned success without retracting anything. The fix was not a better model. It was a dry-run switch, and the publish call now sits inside the branch that the dry run cannot reach. Question four in the business post, can every action be undone in under a minute, was written from that afternoon.
The honest word for it
The honest word for the engine is workflow, and when I explain it to a client that is the word I should be using, because "agent" implies it chooses what to do and it does not. It chooses how. The steps, the order, the gates and the exit are a file I wrote, and when the run does something I did not want, the fix is a line in that file, not a conversation with the model. The one thing on my desk that behaves most like the vendors' definition of an agent is me, reading the log the next morning and deciding whether the file needs a new line.
Chat soon.