Boris Cherny, who runs Claude Code, advised Enterprise Insider in Might that the phrase “vibe coding” had began to bother him, and that he’d gone searching for a greater one. He’s not the one one who’s irritated.
The time period itself doesn’t truly annoy me, although. I believe vibe coding is a very good identify: It describes a particular method of utilizing AI instruments, and in improvement work, names that imply one thing particular are vital. What annoys me is when individuals confuse vibe coding, deliberately or in any other case, with any form of work the place you write code with AI. That confusion factors to a deeper downside: We’ve been utilizing a number of totally different names for lots of various issues, and we aren’t at all times exact about which is which. I believe we have to repair that, and that’s what this text is about: making the case that the identify we’re searching for is “AI-driven improvement” (or AIDD).
The case for this identify comes from the acquainted “X-driven improvement” sample, as a result of I believe it actually matches right here. Software program engineering already has a sample for naming methods of working it takes significantly: test-driven improvement, behavior-driven improvement, domain-driven design. The identify tells you what the work is organized round, and the suffix carries an expectation together with it: There’s a self-discipline hooked up, with requirements, not only a fashion. Put “AI” in that slot and the identify does the identical job. AI-driven improvement says that constructing software program has reorganized itself round AI, and it says it within the vocabulary we already use for the disciplines we maintain ourselves to. It places this fashion of working in the identical household as test-driven and behavior-driven improvement, and that’s precisely the corporate it must be conserving.
Actually, AI-driven improvement is a reputation that’s been sitting in plain sight, and I’ve been utilizing it in my very own writing for some time. It covers every part we do after we construct software program with AI, and I do imply every part. Vibe coding is only one a part of how we work with AI to construct software program. There’s additionally determining what to construct, writing it down, checking what comes again, and standing behind what ships, and AI is in the course of all of that now. No matter we name this fashion of working, it has to cowl the event, not simply the coding. Now, I’m clearly not a impartial celebration right here, however I additionally don’t actually have something to realize; naming is admittedly vital, and I believe we’d like a great identify for what it’s that we’re doing.
However I’ll admit up entrance that the identify has an issue baked into it, and I need to cope with that head on. I lately bumped into Addy Osmani at Foo Camp, and ran the AI-driven improvement identify by him. He identified that constructing software program with AI can be a vary of practices that runs from vibe coding at one finish to agentic engineering on the different. That rang true with me straight away. It additionally highlighted the true downside I’m attempting to unravel, as a result of it means I’m proposing one identify for an entire vary of very alternative ways of working. Can one identify actually cowl methods of working that totally different? It took me some time to work that out, and I’ll come again to it on the finish.
I really feel just like the identify AI-driven improvement actually is smart as soon as you may see what’s flawed with the names we’ve obtained, so I’ll begin there.
What’s flawed with the names we’ve obtained?
Earlier than I choose these names aside, it’s value saying why any of this issues. Naming sits on the core of programming: A factor isn’t actual till you may discuss with it, and referring to issues is most of what we do. There’s an previous line, normally credited to the Netscape engineer Phil Karlton, that there are solely two exhausting issues in pc science: cache invalidation and naming issues. It’s caught round for many years as a result of it’s true (nicely, perhaps one or two different exhausting issues have emerged since then, however it’s the thought that counts). We take naming a variable significantly, so we must always take naming our complete self-discipline at the very least as significantly, as a result of a poorly chosen identify sticks.
So let me take the names we’ve been utilizing one after the other: what each truly names, what it will get proper, and what it leaves out.
Vibe coding
Vibe coding is an exploratory, prompt-first strategy to software program improvement the place builders quickly immediate, get code, and iterate. Andrej Karpathy, one of many founders of OpenAI, coined the time period, which I believe is admittedly helpful as a result of it describes the way in which a number of builders first work with AI and code.
Now, let me be clear about one thing: I’m in favor of vibe coding, and I train it as a very efficient—and, extra importantly, inventive!—solution to generate a number of code. However builders who rely completely on vibe coding lose contact with their code as a result of they let the AI make the entire selections: not simply particular technical selections, but additionally in regards to the structure and the general route of the mission. When that occurs, they typically find yourself constructing one thing that isn’t fairly what they meant. When you must create a product that should do a very particular factor (which describes {most professional} software program improvement), relying solely on vibe coding can go away you with a product that doesn’t truly meet its necessities. That’s a part of the explanation I developed the Sens-AI Framework, which teaches builders when to shift their strategy away from vibe coding, step again to do extra analysis, and apply extra crucial pondering to what the AI is producing.
That is the place the confusion I opened with does its injury (and I’m unsure whether or not it’s what bothered Cherny): When vibe coding will get used because the identify for the entire job, builders will typically assume that it’s completely nice to belief the AI to take over, and that no matter comes out of the AI is the tip of the mission. In different phrases, the identify units the bar: If the work is simply vibes, then vibes are ok, and “ok” is how you find yourself with a pile of code no one truly checked earlier than delivery. So I take into account vibe coding a helpful method, however it falls brief as a whole method of working.
Vibe coding additionally has a built-in restrict, and I realized it the way in which most classes stick, by getting burned. AI is superb at writing code that appears proper and isn’t. I as soon as vibe-coded somewhat bus-tracker app for the B69 close to me in Park Slope (I advised that story in “AI Code Assessment Solely Catches Half of Your Bugs”), and it labored on the primary strive, besides the AI had picked the flawed cease ID and I sat there watching it predict a bus going the wrong way. The code was right. It did the flawed factor. Vibe coding obtained me a working app in minutes, and it had nothing to say about whether or not the app was proper. That half was on me.
Immediate engineering and loop engineering
These two names belong in the identical part as a result of one principally grew out of the opposite. They describe the identical job, getting the precise work out of the mannequin, at two very totally different scales.
Immediate engineering got here first, and for some time it was a really huge deal. It was seen because the core AI ability, and greater than that, it even turned its personal job title: Firms posted prompt-engineer roles with eye-popping salaries, coaching programs appeared in every single place, and loads of individuals reoriented their careers round it. The premise made sense as a result of the way you ask an AI for one thing modifications what you get again. And particularly for individuals utilizing AI to generate code, whenever you ask for code in a obscure method, you don’t get obscure code: you get code that does the flawed factor, as a result of the AI fills in each clean you left, and it’s unlikely to fill all of them in the way in which you meant. That isn’t hallucination. It’s the AI producing precisely what we requested it to. Give the mannequin context about your mission, constraints it has to respect, and a transparent description of the conduct you want, and also you get one thing you may truly use. Immediate engineering is the identify for doing all of that intentionally.
However whereas immediate engineering is an actual ability, persons are now not enamored with the identify, exactly due to the mode of labor that it implies: To most individuals, engineering a immediate means doing one request at a time. When the AI responds to the immediate, you consider the response and write the following one. That one-request-at-a-time fashion is precisely what’s altering about the entire method we work together with AI, and it’s most likely why many AI engineers have grown to dislike the time period. Peter Steinberger, the PSPDFKit founder who went on to construct the open supply agent OpenClaw, posted a line that traveled quick: You shouldn’t be prompting your coding brokers anymore, you have to be designing loops that immediate your brokers. That was a shot straight at immediate engineering.
What’s pushing builders previous one-request-at-a-time prompting is the sheer variety of brokers they will now run. A couple of month after complaining in regards to the time period “vibe coding,” Cherny advised Fortune that he doesn’t write code by hand anymore, and that on a busy day he’s directing hundreds of brokers, or tens of hundreds, directly. You may’t kind prompts quick sufficient to direct ten thousand brokers.
Loop engineering is the identify Addy Osmani gave the brand new ability that Cherny and Steinberger had been pointing at: He wrote up the sample and gave it an actual structure. As a substitute of typing every instruction your self, you construct the system that produces the directions: a loop that dispatches work to your brokers, checks what comes again, and feeds them the following job time and again, with out you in the course of each alternate. The connection between the 2 names is easy. Loop engineering is immediate engineering at scale; the prompts don’t go away, they only cease being typed by you. It’s tempting to oversell that as a result of a well-built loop actually does run with little or no human intervention. However someone nonetheless has to determine what “proper” appears like, and the loop can’t do this half.
I believe loop engineering is an effective identify and an correct one. Designing the loop that drives the agent is an actual ability, and we’d like a phrase for it. However it names the equipment, and equipment has a failure mode: Put an AI agent in a loop with nothing in it that may inform it no, and it generates, checks its personal work, decides the work is nice, and generates extra. There’s no exterior sign, so it finally ends up agreeing with itself on repeat. A well-designed loop makes brokers productive. It could’t let you know whether or not all that equipment seems working software program or one other assured pile of slop, and I need a identify that covers that half too.
Agentic engineering
Cherny stated that he requested Claude for a substitute for “vibe coding” and obtained “agentic engineering,” and whereas that didn’t settle the difficulty, it was an attention-grabbing response from Claude. The time period didn’t come from Claude, although: Andrej Karpathy had coined it a couple of months earlier, virtually precisely a yr after he coined vibe coding, when he declared his personal earlier time period out of date. That’s how briskly these names are shifting. The man who named vibe coding has already changed it.
Agentic engineering is an correct identify for what it describes: you’re not writing the code your self, you’re directing the brokers that do. It’s additionally a little bit of a mouthful, and it isn’t instantly apparent to somebody who doesn’t already know what it refers to. Numerous individuals have advised me they don’t notably prefer it. I discover it completely nice, and it does a strong job of describing that form of work. You could possibly even argue that loop engineering is a type of agentic engineering, and that immediate engineering is technically a less complicated type of it. However vibe coding actually isn’t, as a result of it’s not engineering in any respect. That’s another reason I believe we’d like an umbrella identify that’s pleasant, descriptive, and simply recognizable.
The time period additionally factors at one thing actual about the place this work is heading: Agentic engineering is popping engineers into managers.
A few years in the past I labored for a supervisor who didn’t care, in any respect, in regards to the high quality of the code we shipped. He wished it out the door the second it seemed even remotely viable, and he was infamous for telling us to cease testing and ship. He used to ask why we needed to wait two weeks for the testers to complete, and I’d inform him it takes time to check code. Then he’d ask whether or not we may simply lower a number of the checks, and I’d ask him, “Which a part of the software program are you okay delivery damaged?”
That angle got here again to chew us greater than as soon as. One time we despatched a whole function out to the consumer principally untested, and a bug went straight to customers. The identical supervisor who saved telling us to skip the testing then known as a protracted, depressing assembly to demand to know why a bug had gotten out. I’ll spare you the total drama, which principally got here all the way down to a QA lead getting pressured to lie about what occurred and pin it again on the event workforce. He didn’t care about high quality, however he cared enormously about ensuring the blame for a high quality downside landed on somebody who wasn’t him.
The explanation I’m telling a narrative that occurred years earlier than AI may write a line of code is the blame. The vital a part of that story, and the explanation it belongs on this article, is how accountability obtained managed: My supervisor’s complete system trusted having somebody to pin a high quality downside on. Directing brokers places you in that supervisor’s place, answerable for a workforce’s output, besides the blame-shifting transfer is gone.
It’s actually tempting to consider a fleet of AI brokers as your workforce. You may even give one among them the QA lead function. However when a damaged function goes out, you may’t blame the QA agent, as a result of “nicely, the AI screwed up” isn’t accessible to you: You’re answerable for the AI. You determined how a lot checking the work obtained earlier than it went out, and the consumer with the damaged function isn’t going to simply accept “the AI wrote that half” as a solution, any greater than pinning our untested function on a QA lead mounted something for our customers. Cherny can handle tens of hundreds of brokers, however he can’t hand the duty for what they ship all the way down to the brokers, as a result of an agent can’t maintain it. Directing a swarm is a administration job, and a supervisor owns the workforce’s output. The accountability doesn’t switch, as a result of on the finish of the road there’s nobody left to switch it to.
Blame is value dwelling on, as a result of accountability is the a part of this work that no identify on the vary captures. The loop-and-agent mannequin works, however it solely works with someone making selections about what proper is. Agentic engineering describes the brokers and the engineering simply nice, however someone nonetheless has to personal what the brokers ship, and that’s the half I need the umbrella identify to hold.
Spec-driven improvement
There’s yet another identify I need to cowl, and it’s the one with the oldest roots: spec-driven improvement. The identify means just about what it says: You begin by writing a spec, an outline of what the software program must do, together with issues like acceptance standards and checks, and the work isn’t finished till the code truly does what the spec says. It comes from the identical household as test-driven and behavior-driven improvement, the place you write the checks first and the code has to make them go.
Spec-driven improvement obtained a severe promotion when AI made producing code practically free (though in case you’re a CIO gazing your token invoice, you may disagree, presumably with some extraordinarily salty language). When code is reasonable to generate, most of the price of constructing software program strikes to checking whether or not what obtained generated is correct. The AI fills the generate step, the verification decides what survives, and a human owns the verification.
It additionally picks up the place immediate engineering leaves off. Some time again I wrote that immediate engineering is admittedly necessities engineering, as a result of a great immediate is generally a transparent description of what the software program has to do. Spec-driven improvement is the place that concept was at all times headed: Write the requirement down earlier than the AI generates, and the work has a normal to fulfill from the beginning.
In order that’s the entire vary, and each identify on it’s doing trustworthy work. Whether or not AI-driven improvement is an effective identify for all of it comes down as to whether it’s describing one thing actual: an precise self-discipline, with precise practices, and an individual who’s on the hook for the end result. The remainder of this text is about that self-discipline.
What all these approaches appear to be in apply
So how do these approaches truly play out whenever you’re constructing one thing actual? For me, wherever the work lands on the vary, it comes down to a couple strikes I maintain coming again to.
Write the spec or the contract earlier than the era, not after. When the agent has one thing concrete to fulfill, acceptance standards, a typed interface, a failing take a look at, the work has a normal to fulfill. When it doesn’t, the AI decides for itself what finished appears like.
Put a second opinion within the course of. I run code evaluate throughout a number of fashions, as a result of they fail otherwise, and a discovering one mannequin is certain about is usually one the others missed completely. A reviewer offers the work one thing that may say no.
Give your defects a shared vocabulary. The High quality Playbook leans on the distinction between code that’s flawed in opposition to the spec, code that’s right however does the flawed factor, and conduct no one specified in any respect. These are totally different failures with totally different fixes, and you’ll’t confirm in opposition to a normal you may’t identify. That is previous quality-engineering floor, and I’ve written sufficient in regards to the software program disaster and making use of high quality engineering to AI coding that I’m on board with taking previous concepts and bringing them again. Among the finest of these previous concepts comes from Joseph Juran, one of many founders of high quality engineering: High quality runs in a series from what the person wants all the way in which to what the product does, and each hyperlink in that chain is a spot verification has to occur.
And maintain a human within the judgment seat. The Sens-AI habits I’ve written about are principally about fault-finding: taking a look at what the AI produced and asking what’s flawed with it, taking place a stage after which one other to search out the foundation, as an alternative of trusting it as a result of it ran. That behavior is the a part of the self-discipline solely an individual can provide, and it’s the toughest half to automate, which is why it issues most.
Skip all of that and also you get the factor that’s giving open supply maintainers in every single place heartburn: what the Wall Avenue Journal now calls “vibe slop,” assured, finished-looking output with nothing beneath it. Slop is precisely what era produces when nothing within the course of can push again.
However isn’t there a contradiction right here?
Now I can come again to the query I left hanging initially: Can one identify actually cowl methods of working that totally different? AI-driven improvement is an umbrella time period, and any identify that broad comes with a requirement it has to fulfill earlier than individuals will settle for it, as a result of a reputation that blindly covers every part names nothing. A reputation that really covers every part is one other matter. I sat with that requirement for some time, as a result of it’s actual, and since the precise names don’t face it. Vibe coding names a method of working. Loop engineering names one other. An umbrella over each of them, plus every part in between, had higher be capable to say what stays the identical beneath it.
What stays the identical is that someone owns the end result. After I vibe-coded my bus tracker, no one was going to catch that flawed cease ID however me. When Cherny directs tens of hundreds of brokers, no one owns what they ship however him. The verification modifications with the stakes. A throwaway prototype will get my eyeballs and a shrug, and manufacturing code will get specs, opinions, defect taxonomies, the entire quality-engineering playbook I maintain writing about. How a lot checking the work wants is a choice you make time and again, mission by mission, generally hour by hour. Who stands behind the work just isn’t a choice you get to make. It’s there at each level on the vary.
Take a look at how a lot of that vary the names we have already got cowl, and what each truly names:
- Vibe coding names the exploratory finish of the vary: immediate, get code, iterate, and keep unfastened on goal.
- Immediate engineering names a ability: writing the instruction that will get the precise work out of the mannequin.
- Loop engineering names the equipment: designing the system that feeds these directions to your brokers and retains them producing.
- Agentic engineering names the structure: the fleets of brokers doing the labor, at no matter scale you may handle.
- Spec-driven improvement, with test-driven and behavior-driven improvement behind it, names the verification half of the job: the usual the work has to fulfill earlier than anybody stands behind it.
Each a type of is actual, and each one among them names a chunk of the work. What none of them names is the entire thing the items add as much as, and that’s the job AI-driven improvement does: It’s the umbrella over all 5. The identify doesn’t choose a spot on the vary; it names the factor that’s true in every single place on it: the AI generates, and a human owns the end result.
That’s additionally what makes the identify prone to final (assuming, in fact, that I’m in a position to persuade individuals to start out utilizing it, which I hope I can, as a result of I believe it’s a great time period). Vibe coding, loop engineering, and agentic engineering all describe how this works proper now, and the equipment is altering month-to-month. A number of the items below the umbrella will get changed, and the brand new items will get names of their very own. The umbrella gained’t have to vary after they do, as a result of the factor it names isn’t the equipment. The “-driven improvement” names have already proven they age nicely: test-driven improvement has meant the identical factor for greater than twenty years.
Agentic engineering is actual, and so is loop engineering; in case you’re directing brokers, study them each. Vibe coding is actual too, and I’ll maintain instructing it. AI-driven improvement is the identify for the entire thing, and it earns its “-driven” the identical method test-driven and behavior-driven improvement did: there’s a self-discipline hooked up, and someone owns the end result. AI made producing code virtually free. It didn’t make being answerable for the code free, and being answerable for it’s nonetheless the job.
