Stop Calling Everything an 'AI Agent.' Most of Them Are Just Chatbots With a For-Loop.

🤖 This article was AI-generated. Sources listed below.

The Word 'Agent' Has Become Meaningless — And That's Dangerous

Sometime in mid-2025, the AI industry collectively decided that "chatbot" was passé. Overnight, every product became an "AI agent." Your email summarizer? Agent. Your calendar scheduler? Agent. That thing that generates a Slack message from a bullet list? Believe it or not — agent.

By early 2026, the term has been stretched so thin it could wrap around the entire hype cycle twice. And I'm here to argue: this isn't just annoying marketing — it's actively harmful to the field, to consumers, and to the genuinely groundbreaking agent research that deserves the spotlight.

What an AI Agent Actually Is (Spoiler: Most Products Don't Qualify)

Let's establish a baseline. Researchers at major labs have converged on a relatively clear definition. An AI agent should be able to:

Plan autonomously — decompose a goal into sub-tasks without hand-holding
Use tools — call APIs, browse the web, write and execute code
Maintain persistent memory — track state across a multi-step workflow
Self-correct — recognize when a step failed and adapt its approach
Operate with minimal human intervention — the human sets the goal, the agent figures out the path

Andrew Ng, one of the most respected voices in AI, laid this out clearly:

"An agentic system can autonomously decide on the steps needed to accomplish a goal, use tools to carry them out, and iterate on its own output." — Andrew Ng, Founder of DeepLearning.AI [¹]

Now compare that to what most "AI agents" on the market actually do: they take a single input, run it through one LLM call (maybe two if they're feeling spicy), and return an output. That's not an agent. That's a function call with a landing page.

The Numbers Tell the Story

The agent gold rush is real and measurable. A Stanford HAI report found that when they tested 16 commercially available products marketed as "agents," only four could reliably complete multi-step tasks without human intervention more than 50% of the time [²].

Let that sink in. 75% of self-described AI agents fail the basic agent test more often than they pass it.

The gap between marketing and reality is a canyon, and consumers are standing at the edge wondering why the "revolutionary AI agent" they're paying $49/month for keeps asking them to manually paste in the data it was supposed to retrieve on its own.

Why This Matters More Than Semantics

You might think I'm being pedantic. "Who cares what we call them, as long as the product works?" Here's why I care:

1. Hype cycles have consequences. We've seen this movie before. Remember when every database was "AI-powered" in 2018? Remember the chatbot crash of 2017 when Facebook Messenger bots were supposed to replace apps? The pattern is brutally predictable: overpromise → underdeliver → backlash → funding winter. Gartner's latest Hype Cycle report placed "AI agents" squarely at the Peak of Inflated Expectations, projecting a trough within 12–18 months [³].

2. It drowns out the real breakthroughs. There are genuinely impressive agent systems emerging. Anthropic's Claude has demonstrated increasingly robust tool use and multi-step reasoning. OpenAI's internal research on agents that can operate computers autonomously — clicking, typing, navigating — represents a real paradigm shift [⁴]. Google DeepMind's Gemini-based agents have shown remarkable planning capabilities in complex environments. But when every SaaS product with an API wrapper calls itself an agent, the signal gets buried under an avalanche of noise.

3. It erodes trust at the worst possible time. Public trust in AI is already fragile. A 2026 Edelman survey found that only 37% of Americans say they trust AI companies to be honest about what their products can do [⁵]. Every time someone buys an "AI agent" that turns out to be a slightly fancier chatbot, that trust erodes further. And we need that trust for the moments when AI agents can deliver transformative value — in healthcare coordination, scientific research, and accessibility.

The Counterargument (And Why It's Half Right)

Fair pushback: language evolves, and rigid definitions can stifle innovation. Some argue that "agent" is simply becoming a spectrum — that even a simple tool-calling chatbot sits on the low end of agentic behavior, and we shouldn't gatekeep the term.

There's something to this. Lilian Weng, head of safety systems at OpenAI, has written thoughtfully about "agentic patterns" as a continuum rather than a binary [⁶]. And it's true that drawing bright lines around fuzzy concepts can be counterproductive.

But here's where I push back: a spectrum still needs anchors. If "agent" means everything from a single API call to a fully autonomous system that can plan, execute, and self-correct over hours of independent operation, then the word communicates nothing. It becomes pure marketing vapor. And in an industry where billions of dollars in investment decisions hinge on category labels, that ambiguity isn't innocent — it's strategic obfuscation.

What We Should Do About It

I'm not suggesting we form the Agent Naming Police (though I would absolutely volunteer). But the industry could take some practical steps:

Adopt tiered terminology. Something like "assisted workflow" → "semi-autonomous agent" → "fully autonomous agent" would let companies market honestly while still riding the agentic wave. The EU AI Act's risk-based tiering offers a loose model [⁷].
Benchmark transparency. If you call your product an agent, publish its success rate on standard multi-step benchmarks. Projects like AgentBench and SWE-bench already exist for this purpose [⁸]. Use them.
Investors: start asking harder questions. When a pitch deck says "AI agent," the next slide should show autonomous task completion rates, not just demo videos where everything magically works on the first try.
Consumers: demand receipts. Before paying for an "AI agent," ask: can it complete a multi-step task I've never explicitly shown it how to do? If the answer is no, you're buying a chatbot. That might still be useful! But know what you're buying.

The Bottom Line

The most exciting frontier in AI right now is the development of systems that can genuinely think, plan, act, and adapt without constant human supervision. That frontier deserves a word that means something. Right now, "agent" is being strip-mined for marketing value faster than anyone can replenish its meaning.

The companies building real agents should be furious. The consumers paying for fake ones should be informed. And the rest of us should stop nodding along when someone pitches us a chatbot with a cron job and calls it the future of work.

Call it what it is. Or don't call it at all.