In an interview in early June, Boris Cherny said he doesn’t prompt Claude anymore. “My job is to write loops.” On June 7 Peter Steinberger posted: “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.” Business Insider had the headline up within two weeks of that post: forget prompt engineering, loop engineering is all the rage now. By August, the Udemy courses were there.
Monthly reminder. That might be the most honest thing anyone in this industry has said in three years. The way we’re supposed to work now ships on a release cadence.
I already wrote about the people who sell these futures and never pay for the ones that don’t arrive. This piece is about everyone else: the companies that turn each new word into policy, the dashboards built to measure it, the vendors who relabel their products overnight, the schools that print certificates for it. And about how all of us together forget the last word before the next one lands.
The obituaries
Start in January 2023. Andrej Karpathy writes that the hottest new programming language is English. A few weeks later Anthropic posts a prompt engineer opening with a salary range topping out at $335,000, and TIME runs it as the AI job that doesn’t need a computer engineering background. Vanderbilt’s prompt engineering course becomes Coursera’s top generative AI course of 2023. By now more than 740,000 people have enrolled. Indeed searches for the title go from 2 per million to 144 per million in three months.
In April 2025 the Wall Street Journal calls the job already obsolete. Microsoft’s own survey, as the Journal reports, ranks it second to last among roles companies plan to add. Searches settle at 20 to 30 per million. The promised profession faded. The course is still enrolling.
Spring 2023 also gave us AutoGPT, “an experimental open-source attempt to make GPT-4 fully autonomous.” It collected 100,000 GitHub stars in weeks and became notorious for getting stuck in loops. It’s a low-code workflow builder now.
That March, OpenAI shipped ChatGPT plugins, and they were widely called the App Store moment. Plugins were gone in thirteen months. Custom GPTs replaced them, and by the time the GPT Store opened, users had already built three million of them. OpenAI is scheduled to retire custom GPTs on December 11. The thing replacing them is called plugins.
Pinecone raised $100 million at a $750 million valuation after saying it had introduced the vector database category. Two years later vector search was available inside Postgres through pgvector and across the major cloud databases, and Pinecone was reportedly exploring a sale amid rising competition. Humane raised more than $230 million to replace your phone, sold its assets to HP for $116 million, and the pins stopped working. Replacement, absorption, collapse: different failures, same attention cycle.
Then the cycle sped up. Karpathy coined vibe coding in February 2025: give in to the vibes and forget the code exists. Collins made it Word of the Year in November. In February 2026 Karpathy drew a line between vibe coding for experiments and professional work with agents, and said his favorite name for the second was agentic engineering. The New Stack wrote the obituary under the headline “Vibe coding is passé.” A distinction in one engineer’s post became the end of an era in a headline. Twelve months from coinage to obituary, three from the dictionary to obituary. Context engineering became the phrase everyone repeated in June 2025, pitched as the grown-up replacement for prompt engineering. A year later, a reply under Steinberger’s post announced that harness engineering is so last year.
Loop engineering didn’t get a year. Addy Osmani published an essay titled “Loop Engineering” on June 7. On July 18 Steinberger asked whether we were still talking loops or had moved on to graphs, and a few hours later Hamel Husain posted “Loop Engineering Is Dead. Enter Graph Engineering.” Both were at least half joking. It didn’t matter. Within days the timeline had produced graph engineering roadmaps and tool stacks, and the courses followed. On July 22 LangChain published a retrospective titled “3 Years of Graph Engineering with LangGraph.” Six weeks from naming to obituary, and the obituary was a joke the machine took literally.
The machine
None of this needs anyone to lie. It needs a machine, and the machine has far more parts than a CEO with a microphone. No single actor runs it. Each part answers to the incentives of the part before it.
It starts with a word. Someone with an audience names a practice, and the name is catchy enough to repeat.
Then the pressure that makes a word travel becomes policy. In a memo made public in April 2025, Tobi Lütke made “reflexive AI usage“ a baseline expectation at Shopify and wrote it into performance reviews. Coinbase gave its engineers a week to onboard to Copilot and Cursor, then fired the ones who couldn’t give a good reason why they hadn’t. On the same podcast where Brian Armstrong told that story, John Collison pointed out that nobody really knows how to run an AI-written codebase. Armstrong said “I agree.”
Policy needs a number. The industry measured the one thing it could see. Meta made “AI-driven impact” a core expectation for 2026. I wrote this summer that Meta grades its engineers on token consumption, and that was too blunt. On September 2 Meta told its engineers that AI adoption dashboards and token counts would not be used to evaluate impact, and its review guidance swapped references to AI usage for outcomes that “can be supported by AI or other means.” A spokesperson called it a clarification and said contributions had always been the basis of evaluation. Whatever the policy formally was, the pressure arrived as a core expectation and was later recast as a clarification.
The vendors move next, faster than anyone. In June 2025 Gartner estimated that of the thousands of companies selling agentic AI, about 130 offered anything that deserved the label. Many of the rest had relabeled chatbots and RPA scripts. That year Gartner put AI agents at the peak of inflated expectations. In 2026 agentic AI got a hype cycle of its own, and the category is still at the peak.
Schools arrive last and stay longest. In August, six months after The New Stack’s obituary for vibe coding, Google added a new course on vibe coding to its AI Professional Certificate, which Google calls the most popular generative AI certificate on Coursera. The University of Denver’s five-week vibe coding certificate opened a new cohort on October 5. Loop engineering was declared dead in July, and a Udemy course last updated in August still teaches it as the top of a four-layer stack: prompt, context, harness, loop. Every dead word, laminated into a syllabus. Another loop course promises you’ll be fluent before it becomes a hiring prerequisite.
Then comes the forgetting. I’ve covered Klarna’s round trip from 700 replaced agents back to human support. Even on the way back, a spokesperson described the company as very much AI-first. Seventeen months after his memo went public, Lütke went on a podcast and described what Shopify calls slop grenades: AI output nobody read, tossed to a coworker to check. The way lazy work fails now, he said, isn’t too little output. It’s too much. Nobody retracted the memo. Nobody needed to. The next word had already arrived.
We write postmortems for a six-hour outage. Nobody writes one for a year of mandates.
Why it sticks to AI
Crypto had this rhythm too: ICOs, DeFi summer, NFTs, DAOs, the metaverse, play-to-earn. Most of those never became what they were sold as either. What differs is what the words stuck to.
Crypto words stuck to speculation, and most engineers could smell it. You could ignore NFTs and your job stayed the same. AI words stick to something that works in your editor every day, and sometimes the editor makes the choice for you. Cursor now opens in its Agents Window by default. The setting that’s supposed to turn it off wasn’t in my settings at all, and its forum is full of people for whom it’s ignored or vanished. Cursor’s support answered on its forum that starting there by default isn’t forcing anything. I deleted Cursor last month over a different silent default. Each ephemeral term borrows credibility from a real tool, which makes hype and genuine shift hard to tell apart, and for a manager under pressure the safe move is to adopt the word.
And crypto sold you an asset. This sells you a new version of yourself. Prompt engineer, vibe coder, context engineer, loop engineer, graph engineer: each is a job identity with an expiry date, and each rename arrives with a quiet hint that whoever still answers to the old one is falling behind. The reskilling bill lands on the engineer every cycle, whether the new word survives or not.
The other side
The honest counter is a good one: the labels die, the practices mostly don’t.
Prompting became something everyone does, a capability inside a job instead of a job title. Vector search got absorbed into the database you already run. MCP was pitched as a USB-C port for AI and actually became a standard, now under the Linux Foundation’s Agentic AI Foundation. Lovable lost about 40% of its traffic by September 2025, according to Barclays. By June 2026 the company reported roughly half a billion dollars in annualized revenue. Coding agents are real. Most of my own code is written by one.
Some renames track real change, too. Karpathy’s stated reason for moving past vibe coding was that the models got better. And naming churn is older than LLMs. Ask anyone who watched operations spawn DevOps, SRE and platform engineering.
All fair. But if the practice survives and only the label changes, the honest name for it is a version bump, and version bumps don’t usually arrive with leaderboards and five-week university certificates. Not every mandate follows a rename. The shared habit is making adoption compulsory before the previous promise has been measured. Label after label is sold as a new era, with mandates, metrics or syllabi attached, then swapped for the next one before anyone checks what the last one did. And the successors keep arriving faster: roughly two years from prompt engineer’s peak to its obituary, one year for vibe coding, less than a year before context engineering had a successor of its own, six weeks for loop engineering.
What survives every rename
Look at what a loop actually is. You hand the model a goal, let it run, check the result against something, and repeat until the check passes. That was AutoGPT in 2023. What changed since is better models and loops built into the tools. That’s real progress. It isn’t a new profession. Neither is the graph, this summer’s successor: nodes, edges and state connecting loops, branches and checks, the kind of control flow workflow engines have drawn for years. One reply to Husain’s post said simply “welcome back, langchain.”
The people building loops say the important part out loud. Addy Osmani: “The model that wrote the code is way too nice grading its own homework.” Paweł Huryn: “the loop is the easy part.” The real work, he says, is the stop condition: the check that ends it, the budget that caps it, a target the system can actually reach. Every layer of that stack, graphs included, ends in the same place, with someone deciding what done means and recognizing output that only looks done.
Each rename moves the engineer one layer further from the code and acts as if the layer below were solved. The one thing it never makes obsolete is the judgment at the top, and that is much harder to package as a new era. Five weeks of coursework won’t hand it to you. You build it by doing the work the last word told you to stop doing.
Right now a site called FindSkill sells a Professional Certificate in Loop Engineering. The term’s obituary is almost three months old, the graph engineering courses are already on sale, and whoever holds that certificate will still need the one thing no certificate can vouch for: knowing when the output is wrong.