Say the word "agentic" out loud ten times.

By the fifth repetition it starts sounding like a medication side effect. By the tenth, you've experienced what psychologists call semantic satiation: the cognitive phenomenon where repetition drains a word of all meaning until it becomes a hollow collection of phonemes rattling around your skull.

The AI industry managed to do this at industrial scale, to an entire vocabulary, in under eighteen months.

The Agenticpocalypse#

I counted the word "agentic" in various reports and press releases for major companies. I stopped at 400 because I ran out of will to live.

Salesforce has an "agentic" CRM. ServiceNow has an "agentic" workflow platform. Microsoft bolted "agentic" onto Copilot so many times that the word appears more frequently in their marketing copy than "the." Bosch published an entire whitepaper about "agentic AI" for HVAC systems. Your heating and cooling is now agentic. It adjusts the temperature. That's it. That's the agent.

The word has been stretched so thin it could be used as cling film. An "agent," in any meaningful sense, implies autonomy, goal-directed behaviour, the ability to plan and execute multi-step tasks without a human holding its hand. What we've actually got is a timer attached to an if-statement.

Somewhere a product manager is reading this and preparing to describe their next feature update as "agentically enhanced." I can feel it.

A Glossary of the Deceased#

"Agentic" isn't alone in this mass grave. The entire AI lexicon has been through the linguistic equivalent of a car crusher.

"Reasoning." This used to mean something. Deductive logic. Chain-of-thought problem solving. Drawing valid conclusions from premises. Now it means "the model produced more than three sentences before answering." I've seen demos where a vendors describe their model's "advanced reasoning capabilities." and it turns out to have been a solution where you would get more rigorous logic from a magic eight ball.

"Multimodal." Originally: a system that can process and generate across fundamentally different data types (text, images, audio, video) in an integrated way. Now: "our chatbot accepts file uploads." Congratulations, you've reinvented the email attachment and given it a research paper.

"Hallucination." What started as a precise technical term for models generating fabricated information has become a catch-all excuse. "The model didn't hallucinate, it was being creative." "That's not a hallucination, it's an alternative interpretation." At some point we collectively decided that the polite term for "confidently wrong" is "hallucinating," which does a massive disservice to both the technical phenomenon and anyone who's actually experienced hallucinations.

"Alignment." Originally referred to the deeply important problem of ensuring AI systems behave in accordance with human values and intentions. Now it's used to describe any fine-tuning pass that makes the model slightly less likely to tell you how to hotwire a Honda Civic. The entire field of AI safety got compressed into a content moderation checkbox.

"Guardrails." Speaking of which, every content filter is now a "guardrail." Every rate limiter is a "guardrail." I've seen API timeout settings described as "guardrails." The word has been so thoroughly defanged that it now refers to any code that prevents a product from doing something, which is, by definition, most code.

How Words Die#

There's a well-documented pattern in marketing called the "term lifecycle." A technical concept gets coined by researchers. It filters into industry discourse. Marketing departments identify it as valuable shorthand. They strip it of specificity, broaden its application until it covers everything and therefore describes nothing, and then move on to the next victim.

"Cloud" went through this. "Big Data" went through this. "Machine Learning" went through this so hard that by 2019, a sorting algorithm counted as ML if the sales team squinted hard enough.

But AI terminology is speedrunning the process. "Agentic" went from a useful descriptor in a handful of research papers to a meaningless marketing adjective. The acceleration is partly because the stakes are higher (AI companies are currently valued at GDP-of-small-nations levels, and every press release needs to sound like a breakthrough) but it's mostly because there are now approximately nine thousand AI startups and they all need to differentiate using approximately twelve words.

The result is a vocabulary arms race where everyone uses the same terms to describe wildly different things, and the terms themselves become noise. When everything is "agentic," nothing is.

The Real Damage#

This isn't just a pedantic gripe about marketing copy, though it is also that.

The hollowing-out of technical vocabulary has tangible consequences. When a researcher publishes a paper about genuine advances in autonomous agent architectures, they have to use the same word, "agentic", that Bosch used to describe air conditioning.

Enterprise buyers making million-dollar procurement decisions can't distinguish between products because they all describe themselves identically. "Our agentic AI platform leverages advanced reasoning with multimodal capabilities and built-in guardrails." That sentence describes approximately 300 products, and I've seen it, near verbatim, on at least forty landing pages.

Worse, the semantic collapse makes it actively harder to talk about actual limitations. If "reasoning" already means "produces text," how do you explain to a non-technical stakeholder that the model doesn't actually reason? You've lost the linguistic tools to make the distinction. The vocabulary gap between what these words mean in a research context and what they mean in a sales call is now so wide you could put the offices of the whole ASX200 in it.

The "AI-Native" Infection#

The latest mutation is the suffix "-native." AI-native. Cloud-native. Agent-native. It's the linguistic equivalent of adding racing stripes: purely cosmetic, zero performance gain.

"AI-native" is supposed to mean "built from the ground up with AI at its core," as opposed to bolting a ChatGPT wrapper onto an existing product. In practice, it means "we put AI in the product name and would very much like more funding please." I've reviewed "AI-native" applications that are, upon inspection, a Flask app making API calls to GPT-5. That's not AI-native. That's API-native, and the distinction matters to everyone except the people writing the cheques.

The suffix has also spawned the magnificently absurd "agentic-native," which I first encountered last month and immediately needed to lie down. We've entered the phase where buzzwords are reproducing with other buzzwords, creating compound shite at a rate that outpaces anyone's ability to call it out.

A Modest Proposal#

I'd like to suggest a moratorium. Not on AI development (that's clearly unstoppable) but on the current vocabulary. Retire the lot of it. Every word that's been ground into meaninglessness gets put out to pasture for a minimum of two years.

In the interim, companies must describe what their product actually does using only words that a plumber would understand. Not a plumber with an interest in AI. Just a plumber.

"Our software reads your emails and writes draft replies." Clear. Honest. Nobody's confused.

"Our platform automates invoice processing by matching line items to purchase orders." Brilliant. I know exactly what it does.

"Our agentic AI-native platform leverages frontier multimodal reasoning to deliver aligned, guardrailed intelligence at scale." This tells me nothing. Absolutely nothing.

The test should be simple: if you remove every buzzword from your product description and there's nothing left, you don't have a product.

What Comes Next#

The frustrating part is that genuinely interesting things are happening in AI right now. Models are getting meaningfully better at sustained multi-step task execution. There are real architectural innovations in how systems maintain state across complex workflows. Progress in tool use and environment interaction is tangible and measurable.

But we can't talk about any of it clearly because the vocabulary has been pre-ruined. Every legitimate advance has to fight through a fog of marketing-speak to get noticed, and by the time it does, the words used to describe it have already been claimed by fourteen SaaS products and a smart fridge.

Semantic satiation is supposed to be a temporary cognitive phenomenon. You say a word too many times, it briefly loses meaning, and then your brain resets. The AI industry's version is permanent. These words aren't coming back. They've been too thoroughly hollowed out.

So we'll need new ones. And when we get them, I give it about six weeks before someone describes a calendar reminder as "cognitively autonomous" and the whole cycle starts again.