Beyond the Prompt: What Agentic AI Actually Means to Me

• The terminology matters
• My move into agentic AI
• Why I chose Hermes
• The tradeoffs

I used to think “agentic AI” mostly meant a smarter chatbot.

That definition is too broad to be useful.

This is the first post in a series about my journey with AI in my IT work. I will share what I have learned from my own experiences and experiments, including where AI helps, where it falls short, and what I am still figuring out. These are my observations, not universal rules.

§The terminology matters

I find it helpful to separate AI terminology into the following patterns:

• Automation follows steps someone designed in advance.

• A chatbot responds to a user.

• A copilot helps a person produce or review work.

• An agent receives a goal, chooses among approved actions, checks what happened, and decides what to do next.

• A multi-agent workflow divides that work across several specialized stages.

The boundaries overlap. A chatbot can use tools. An automation can include an AI step. A copilot can take an action.

Most of us already understand the chatbot experience. You do not have to be an IT professional to open Copilot, ChatGPT, Claude, Gemini, or another AI tool and use it to ask a question, generate an idea, summarize information, or help create something.

That is where most folks stop.

§My move into agentic AI

I recently made the leap into agentic AI, and it changed how I think about the technology. The focus moved from asking AI for an answer to giving it a goal and allowing it to work through the next steps, within boundaries that I define.

One recent example involved work I had done on a wiki for a Power Platform GitHub Center of Excellence. I was asked to present the work, but I did not want to simply show the wiki and walk through it by reading from the page. I used Hermes to review the material and put together a presentation on my behalf. When I came back, I reviewed it for accuracy, added comments and tweaks, and had a working presentation in less than an hour. Creating that presentation manually would normally have taken the best part of a morning or afternoon.

An agent is not a fire-and-forget service. I think of it more like delegating work to a co-worker. I expect to review what it creates, course-correct when necessary, and make the final decisions. That review loop helps avoid generic AI slop and keeps the result grounded in my own thinking and experience. The work may be assisted by an agent, but the outcome is still inherently mine.

§Why I chose Hermes

I chose Hermes as the AI agent platform I wanted to explore. OpenClaw is another example that has received a lot of attention from AI commentators. It is an open-source personal AI agent that can connect to tools and carry out tasks across different systems. For me, though, Hermes became the better fit for learning how to work with agents in a deliberate and controlled way.

Hermes can turn lessons from a workflow into reusable skills, which means I am not starting from scratch every time. It has also given me more confidence in setting security boundaries, reviewing permissions, and deciding when an action should require my approval.

I use Hermes as a working partner for developing ideas, researching and explaining technical concepts, creating and editing files, and breaking larger IT tasks into manageable steps. I am still the person setting the objective, deciding what information is appropriate to use, reviewing the result, and determining what should happen next.

I run Hermes on an always-on Virtual Private Server (VPS), but the architecture behind that setup is a topic for another day. The important point for me was not having an agent running directly on my daily-driver PC or Mac with unfettered access to my important files, applications, and data. Separating the agent from my primary devices made it easier for me to think deliberately about access, permissions, and what the agent should be allowed to do.

§The tradeoffs

Here is my two cents on where agents shine and where they do not.

The biggest advantage for me is that work can continue during my down time. I can frame a project around outcomes, deliverables, and scope, then let the agent handle much of the heavy research and development work while I am doing something else.

That also changes my role. Prompting is usually a question-and-answer interaction. With an agent, my role shifts more toward delegating: defining the objective, providing context, setting boundaries, reviewing progress, and evaluating the result.

There are tradeoffs. Getting started requires more advanced setup than opening a chat window. Choosing the right model matters, both for the quality of the results and for managing cost. Security is paramount, especially when deciding how the agent is sandboxed and what systems or data it can access. Sharing memory and context between agents also requires planning. I will save that topic for another day.

A loop is not automatically an agent. A loop with broad permissions and no useful audit trail is a problem waiting to happen.

My working definition is:

Agentic AI is a workflow pattern where a model can pursue a goal through approved tools, decisions, feedback, and controls.

Are you using agentic AI in your work? Has it helped you get more done, or has it introduced new friction and risks?