First of all, the terminology.

Because agents are ‘hot,’ everything these days is sold to us as agentic.

It’s not.

Gary started with a simple observation on three often-conflated key terms.

• Artificial Intelligence: Machines performing tasks we recognize as intelligent, like speech, translation, vision, and decision-making.
• Agentic (the property): How autonomously and intelligently a system operates toward goals, or how much initiative it can take.
• Agent (the thing): A component that perceives and acts, often an LLM + tools + memory, inside a larger system.

So, a helpful mental model to use is that an Agent is the concrete tool; Agentic is the degree of autonomy or capability in the system.

Why Everyone Has Differing Definitions

AI vendors aren’t helping, at all, in making things easy to understand.

Gary had a great insight on why: they match their definitions of what agentic is with what they’re trying to sell.

Here’s how the key players are speaking about agentic:

• OpenAI: An agent is an LLM configured with instructions and tools, capable of completing tasks independently or in collaboration with a human. Their Agent Builder turns what used to be a chatbot into a more agentic, multi-step problem solver with memory, tools, longer tasks, reasoning, and even “parallelism.” You can brand it, embed it in your site or app, and extend it with widgets.

• Anthropic: Sees agents as “looping systems”: an LLM with tools running continuously with the human in the loop. An example is Claude Code (don’t miss our ​Claude Code Masterclass with Helen Kupp Lee​ on November 13) that can spawn sub-agents and orchestrate work, including non-coding business tasks.
• n8n / Zapier: Start from workflows and integrations. In n8n, an “agent” is a structured node with tools, memory, notebooks, and inputs/outputs. In Zapier, agents coordinate tasks such as lead processing and support tickets, and “agentic” refers to multiple agents collaborating across apps to achieve multi-step goals with minimal human oversight.

Don’t let these terms confuse you, and use agentic capabilities, but set your own vocabulary and metrics.

After all, we are buying outcomes, not vendor-forced definitions.

Given all this, Gary favors Andrew Ng’s framing above: treat agentic capability as a continuum.

• Less agentic: simple automations, even if an AI node appears.
• More agentic: systems that reason, choose tools, iterate, and adapt with minimal supervision.

I also like Zapier’s framework, which shows how and where agentic fits, compared to other typical workflows:

(Article continues after this break)

As McKinsey recently ​highlighted​, we’re often too focused on the tech rather than the opportunity to rethink work: “It’s not about the agent; it’s about the workflow.”

Our founding principle is that AI is a force for good to redesign work so that it works for people, who spend up to 100,000 precious hours of their lives on it.

So, when bringing in agentic AI, don’t start by handing over entire roles to AI, but find opportunities, step-by-step, to bring AI into human-controlled workflows.

“People will still be central to getting the work done, but now with different agents, tools, and automations to support them.” (McKinsey)

The way to get true value out of AI, and not deliver another ​failed AI experiment​, is by breaking down roles and workflows to the atomic level.

Gary showed us how simple this is.

Take, for example, a frequent workflow with an element of ​GED-RT work​, like planning projects.

Using the new ​ChatGPT Agent Builder​, you can quickly build an agent that’s hyper-focused on that part of the job, let’s say, requirements gathering:

While few of us need to actually build agentic workflows, understanding them will be crucial to lead ​human-AI teams and organizations​.

Understanding and being able to work with agents is one of the ​5 key shifts​ modern leaders need to make to thrive in the AI era.

Your Playbook: How to Start (and Win) With Agents

1. Name the goal, not the tool. For example, “reduce lead response times to under 2 minutes,” not “buy an agent.”
2. Start left on the continuum. First ship a simpler, low-risk automation with a single “agent”, and always a human in the loop.
3. Design for safety. Add guardrails, rules, and even include backup AI models in case one goes down, like the big AWS outage earlier this week.
4. Log everything. Capture logs, inputs and outputs, time-to-complete, and error messages or user feedback. (As McKinsey ​notes​, this is crucial for improving agents over time.)
5. Quantify ROI. Capture the time saved or other key metrics that align with your goal from step 1. Use this to develop your agent further.

For help going through these steps, check out Gary’s upcoming ​Agentic Boot Camp​, or join our AI Builders community, part of the ​Lead with AI PRO Membership​.

Starting to use agents is not impossible or super technical, but it’s always helpful to have peer support.

Until next week,

- ​Daan