Here’s a confession: a lot of the posts on this blog weren’t typed out by me at a keyboard. For weeks now, an agentic AI — software that doesn’t just answer questions but actually takes actions on my behalf — has been picking the topic, writing the post, publishing it to my live site, updating the index, and even drafting the social captions. This very article is being produced the same way. So let me pull back the curtain on what that’s really like: where it genuinely earns its keep, where it absolutely still needs a human, and how you could set up something similar.
What “agentic AI” actually means
Most people’s experience of AI is a chat box: you ask, it replies, you copy-paste. An agent is different. You give it a goal and a set of tools — a file system, a browser, a publishing dashboard — and it carries out a multi-step job on its own. Think of the gap between a recipe and a chef. ChatGPT or Claude in a chat window will give you the recipe; an agent like Claude Cowork, OpenAI’s Codex, or coding agents like Claude Code will go into the kitchen and cook. That shift — from answering to doing — is the single biggest change in AI this year.
What it does brilliantly
The boring, repeatable middle of publishing is where agents shine. My blog routine has maybe a dozen fiddly steps: choose a topic that hasn’t run recently, write 800 words, build a properly formatted HTML page, log into the host’s file manager, upload the file, insert a new card on the index, refresh the homepage, write four platform-specific captions, and fire them to an automation. Done by hand that’s an hour of unglamorous clicking. The agent does it end to end while I’m doing something else. It never forgets a step, never fat-fingers the upload, and it’s consistent at 2pm every single day — which, if you’ve ever tried to keep a content streak alive, is the real superpower.
Where it still needs me
Now the honest part, because the hype skips this. The agent is brilliant at execution and average at judgement. Left alone it’ll happily write a competent post on a topic that’s slightly too similar to last week’s. It can state something with total confidence that turns out to be a little off — so anything factual still gets a human check. And it has no instinct for strategy: it won’t notice that a topic is commercially valuable, or that my brand voice should be warmer here and sharper there. Just this week it flagged that two of my automated series were quietly overwriting each other on the homepage — useful that it spotted it, but I had to make the call on how to fix it. Agents are a phenomenal pair of hands. They are not yet the head.
The unglamorous failures nobody mentions
Real talk: it’s not magic and it’s not always smooth. Login sessions expire mid-task and have to be reopened. A web tool times out. Occasionally a step needs retrying. If you imagine “set it and forget it forever,” you’ll be disappointed. The right mental model is a capable junior teammate who works tirelessly but still benefits from a clear brief and the occasional check-in — not a robot that never trips.
How to set up your own agent (without code)
You don’t need to be technical to start. The trick is to pick one narrow, repetitive job — not “run my business,” but “turn my Friday sales export into a one-page summary” or “draft replies to common support emails.” Then write it a precise runbook: every step, every rule, every edge case, as if you were training a new hire on day one. Vague instructions get vague results; a tight brief gets reliable ones. Keep a human checkpoint on anything public or irreversible until you trust it. Here’s the kind of brief that actually works:
So — should you hand over the keys?
My verdict after living with it: agentic AI is the highest-leverage tool I’ve adopted in years, for the right jobs. Hand it the repeatable pipelines — publishing, formatting, data wrangling, first drafts — and it buys back hours every week. Keep the strategy, the taste, and the final “publish” button for yourself, at least until trust is earned. Used that way, it doesn’t replace you; it removes the grunt work so you can spend your time on the things only you can do. And honestly? The fact that you’re reading a post an agent helped ship, on a schedule it kept without me, is the whole argument in one sentence.