April 10, 2026

AI for the Over 40 – Week 22: When Everything is an Agent, Nothing is an Agent

By the second day of an AI + ERP conference I attended, I had heard the word ‘agent’ so many times that it had lost all meaning.

A researcher is an agent. An analyst is an agent. Facilitator is an agent. Every sponsor demo featured agents. Every hallway conversation mentioned agentic AI.

Then another attendee said what I had been thinking: there don’t seem to be any common definitions of what “agent” or “agentic” actually means.

Everyone nodded. No one clarified. A week earlier, a technology leader told me some of his employees had saved well-crafted prompts as reusable instructions and genuinely believed they had built an “agent.”

That’s when it clicked. The word agent now covers everything from a smart saved prompt to systems that work while you sleep. If everything is an agent, the word means nothing.

And that confusion is going to cost organizations real money.

Why defining ‘agent’ is genuinely hard

My first instinct was to blame marketing. Vendors labeling everything as agent to ride the hype cycle.

But the problem runs deeper. We struggle to define agent because we cannot see what is happening inside these systems.

I call this the observability problem.

What we can see and control

There are external variables we can observe and manipulate.

Model selection. Choosing between models like Claude, GPT, or Gemini dramatically affects speed, depth, and behavior.

Mode selection. Standard chat versus deep research, extended thinking, or agent mode.

Project configuration. The instructions, documents, and context we provide.

Connectors and tools. MCP servers, plugins, and integrations that extend access to systems.

These are the switches we control. We can test them. We can compare results.

What we cannot see

But internal behavior is largely hidden.

• How many reasoning steps did the model take?
• Did it spawn parallel tasks?
• Which internal tools did it call?
• Did it revise its own work before responding?

The same external inputs can produce very different internal processes. And we often have no visibility into why.

The gap that creates confusion

This gap between what we control and what we can observe creates the definition problem.

Is something an agent because of internal complexity? If so, we often cannot tell.

Or is something an agent because of what we experience—how much autonomy we grant and how independently it operates?

Since internal mechanics are opaque, defining agent by what is happening inside is a dead end. We need to define it by experience. By observable autonomy.

An experience-based framework for agent

After dozens of conversations and experiments, I have settled on a four-level spectrum. The framework focuses on what you experience, not what marketing claims.

Level 1: enhanced chat

Reactive. You ask, it responds.

It may reason deeply, search broadly, or take longer to answer. But nothing happens when you close the window.

Examples include deep research modes, knowledge assistants, and advanced analyst features.

Saved prompts that consistently produce strong results live here. Useful? Absolutely. But still enhanced chat.

Level 2: triggered automation with AI reasoning

A predefined trigger starts a predefined workflow.

AI makes decisions within structured paths.

This is workflow automation with an LLM brain.

Examples include event-triggered assistants, meeting summarizers that generate tasks, and prebuilt business agents.

Sophisticated and valuable. But the path is still largely defined in advance.

Level 3: goal-pursuing agent

You provide a goal. The AI plans and executes steps to reach it.

It decides the path within the guardrails you set. It may adapt based on what it discovers. It may ask clarifying questions or request approvals.

But it is session-bound. When you close the window, it stops.

Examples include multi-step research tasks, computer-use capabilities, and complex tool-driven sessions.

The key question is who decides the next step. At Level 2, rules and triggers decide. At Level 3, the AI does.

Level 4: persistent autonomous agent

This is what most people imagine when they hear the word ‘agent’.

It works across time. It may continue while you are away. It sets sub-goals, adapts when things fail, and maintains state across sessions.

Examples include long-running coding agents or research systems that operate overnight.

This is true persistence and autonomy.

What agent-like behavior actually feels like

Recently, I tested ChatGPT’s Agent Mode with a research task. I gave it a goal and walked away.

Eleven minutes later, it was finished. During that time, it searched, evaluated sources, synthesized findings, and structured results. I was not guiding each step. I was delegating.

That felt different. Not because I could see internal mechanics, but because my experience changed. I gave a destination and waited.

Compare that to a deep research response in standard chat. The output may be equally impressive. The internal complexity may be similar. But experientially, I asked a question and received an answer. That is enhanced chat, not agency.

Platform differences reinforce this confusion. Claude offers strong agentic capabilities through Claude Code for developers, but not in its standard chat interface. ChatGPT brings agent-like behavior into chat through Agent Mode. Capabilities vary by vendor and even by product tier. The word agent does not mean the same thing everywhere.

The diagnostic questions

The next time someone claims something is an agent, ask three questions.

Does it work while I am away?
If no, it is Levels 1 through 3. If yes, it approaches Level 4.

Who decides the next step?
If you direct each interaction, it is Level 1.
If triggers and workflows decide, it is Level 2.
If the AI decides within your guardrails, it is Level 3 or 4.

What happens when it fails?
If it stops and waits for you, it is Levels 1 through 3.
If it autonomously tries alternatives, it is Level 4.

These questions cut through buzzwords and force clarity.

From consumer to informed evaluator

Without a framework, you are a consumer of marketing. You hear ‘agent’ and assume autonomy you cannot verify.

With a framework, you become an informed evaluator. You place capabilities on the spectrum. You match what you are getting to what you actually need.

And here is the overlooked truth: sometimes Level 2 is exactly what you need.

If your problem is routing requests or summarizing meetings into structured actions, a persistent autonomous system may be unnecessary. Triggered automation with AI reasoning might be the right tool.

The goal is not to chase Level 4. The goal is to match capability to need.

Why this matters for your organization

Terminology confusion creates real risk.

Misaligned expectations. Leadership expects autonomous agents. You deploy structured workflows. The pilot is labeled a failure.

Wasted investment. You pay premium prices for agentic AI that is effectively enhanced chat.

Missed opportunities. You dismiss Level 2 solutions as unimpressive when they are exactly what your business needs.

Literacy matters. You cannot delegate to something you do not understand. You cannot evaluate what you cannot define.

Your week 22 challenge: cut through the noise

This week, apply the framework.

Find an agent claim in a vendor pitch, product demo, or internal project. Ask the diagnostic questions. Place it on the spectrum. Identify the gap between implication and reality. Then ask whether that level of agency matches your need.

Share what you discover with a colleague. The more leaders who adopt clear definitions, the less money organizations will waste on ambiguous promises.

The bottom line

I left that conference realizing we do not have shared definitions for the term ‘agent’. That confusion will cost organizations money.

We cannot reliably define an agent by internal mechanics because we cannot see them. We must define it by experience. By autonomy granted. By what happens while we are away.

This framework is not about hype. It is about clarity.

If everything is an agent, nothing is an agent. And you cannot invest intelligently in a word that nobody can define.

This post is part of my “AI Over 40” series. It first appeared on LinkedIn: AI for the Over 40 [Week 22]: When Everything Is an Agent, Nothing Is an Agent

Read more AI and Copilot blogs.

Trending Posts