In February 2025, Andrej Karpathy — co-founder of OpenAI and former AI lead at Tesla — posted a casual tweet that became a manifesto. He described a new way of programming he called “vibe coding,” where developers fully give in to the vibes, embrace exponentials, and forget that the code even exists. He barely touched the keyboard, accepted all AI-generated changes without reading them, and copy-pasted error messages back to the model until things worked. His caveat was telling: this approach was “not too bad for throwaway weekend projects.”
One year later, in February 2026, Karpathy declared vibe coding passe. LLMs had gotten smart enough that professionals were using AI agents as a default workflow — not just for throwaway projects, but for real engineering. He now calls it “agentic engineering,” emphasizing that orchestrating AI agents requires genuine expertise.
Between those two posts lies a fundamental question reshaping the software industry: if AI makes code nearly free to generate, should we keep maintaining it at all?
The Economics That Changed Everything
Software has always been expensive because developers are expensive. Building a modest application required skilled engineers, months of work, extensive testing, and long-term maintenance planning. Organizations optimized for longevity because the cost of rebuilding from scratch was prohibitive.
That calculus is breaking down. According to the 2025 Stack Overflow Developer Survey, 84% of developers now use AI coding tools, and AI generates or assists with roughly 41% of all code written globally. GitHub reports that Copilot generates 46% of code for its users, with Java developers reaching 61%. Developers report productivity gains of 25 to 39% on general tasks, and 30 to 55% on scoped work like writing functions, generating tests, or producing boilerplate.
Chris Paik, a venture capitalist who published the influential essay “The End of Software” in 2024, put it bluntly: LLMs have proven themselves remarkably efficient at writing code and will drive the cost of creating software to zero. His prediction: when software no longer has to justify its existence through revenue, we will see a Cambrian explosion of purpose-built, single-use applications.
Anish Acharya, General Partner at Andreessen Horowitz, made this concrete in a 2025 essay titled “Disposable Software.” He described building a math game to help his children earn screen time, and a mini Instagram clone called Catsagram for sharing family cat photos with a small group. Neither would have made economic sense to build before AI. His key insight: software creation used to be constrained by return on investment — now it is constrained only by imagination.
The Bifurcation: Disposable vs. Durable
The most useful framework for understanding this shift comes from Charity Majors, CTO of Honeycomb, the observability platform. In a widely discussed 2025 essay, she argued that software development is bifurcating into two fundamentally different domains: disposable code and durable code. This is not a spectrum — it is a fork, and the two paths require entirely different engineering mindsets, toolchains, and quality standards.
Disposable code includes prototypes, data processing scripts, design mocks, internal tools, automation workflows, and hobby applications. This category has exploded since ChatGPT launched. A developer who once spent days writing a data migration script can now generate one in minutes, run it, and discard it. A product team that needed weeks to build an internal dashboard can now spin one up in an afternoon.
Durable code is different. It runs bank transactions, package deliveries, medical records systems, satellite operations, and airline flight paths. Majors argues this code will remain durable until the stars die out because the stakes are high, the losses are material, and the ripple effects of failure are complex. The cost of software, she notes, is defined by its maintenance — and maintenance is precisely what disposable code avoids and durable code demands.
The difference is not whether AI or a human wrote the code. It is about lifecycle standards, reliability requirements, and what happens when things fail.
Where Disposable Software Already Works
Disposable software is not a theoretical future. It already operates in several well-established patterns.
One-time scripts and data work. Developers have always written throwaway scripts for data migrations, file transformations, and database cleanups. AI simply makes this faster and more accessible. What once required a developer can now be done by an analyst describing the task in plain language.
Internal tools. Companies routinely build dashboards, admin panels, and workflow tools used by a handful of employees. These tools rarely justify long-term maintenance budgets. With AI, regenerating them when requirements change becomes cheaper than maintaining them.
Prototypes and experiments. Startups build prototypes to test hypotheses. The entire point is to learn quickly and discard what does not work. AI makes the build-test-discard cycle dramatically faster, reducing the cost of failed experiments.
Temporary integrations. Businesses frequently need short-term connections between systems — a sync between two APIs during a migration, a reporting bridge during a platform transition. These integrations have natural expiration dates.
Personal and micro-audience tools. This is the category Acharya highlighted at a16z. Before AI, building software for five people was absurd economics. Now someone can build a custom meal planner, a niche hobby tracker, or a family photo app in an afternoon. Use it for a while, discard it, build something new.
Advertisement
The Uncomfortable Counterarguments
The disposable software thesis has real limits, and honest proponents acknowledge them.
AI does not always make developers faster. A 2025 randomized controlled trial by METR — a respected AI evaluation organization — found that experienced open-source developers were actually 19% slower when using AI coding tools on their own repositories. The developers themselves predicted a 24% speedup. The gap between perception and reality is significant. AI tools excel at greenfield generation but can struggle with the accumulated context of mature codebases — precisely where maintenance matters most.
Trust remains low. The 2025 Stack Overflow survey found that 46% of developers do not fully trust AI-generated code, and only 3% express high trust. Disposable code that nobody reads before running creates obvious risks in any environment handling sensitive data or real-world operations.
Technical debt does not disappear — it transforms. Developers already spend 33% of their time dealing with technical debt, according to a Stripe study, and the CISQ estimates that poor software quality costs the US economy $1.52 trillion annually. If disposable code proliferates without discipline, organizations risk replacing old-fashioned technical debt with a new kind: thousands of untracked, unmaintained, AI-generated scripts scattered across the enterprise with no documentation, no tests, and no owner.
Infrastructure costs are real. Generating and running disposable software is not free. Each throwaway application still consumes compute, storage, and network resources. At organizational scale, the cumulative infrastructure footprint of disposable software may offset some of the savings from not maintaining it. There is also an environmental dimension: as data centers already strain global energy grids, a world where software is routinely generated, run, and discarded raises questions about the sustainability of treating computation as disposable too.
Code quality concerns persist. GitClear’s 2025 research found a fourfold increase in code clones — near-duplicate code blocks — in repositories using AI assistants. More generated code does not automatically mean better code. Without review discipline, disposable code risks becoming the junk mail of the software world: cheap to produce, expensive to deal with at scale.
What Actually Changes for Engineers
If the disposable software paradigm takes hold, the developer role shifts rather than shrinks.
Writing code from scratch becomes less important. Evaluating, validating, and orchestrating AI-generated code becomes central. System design, architecture thinking, and understanding business context become the differentiating skills. Karpathy’s evolution from vibe coding to agentic engineering illustrates this perfectly: the leverage comes from AI, but the judgment comes from the engineer.
Charity Majors points to another shift. A new product or feature might start as a disposable prototype, then get converted into durable code once its shape comes into focus. Engineers who understand both modes — who know when to throw code away and when to invest in making it robust — will be the most valuable.
The cultural challenge is real. Software engineering has long celebrated craftsmanship: elegant code that stands the test of time. Disposable software asks engineers to treat code as a temporary tool rather than a permanent artifact. This is the same mental shift the industry went through with cloud infrastructure, where servers went from carefully maintained machines to ephemeral resources spun up and destroyed on demand.
What Stays Permanent
Not everything becomes disposable. Critical infrastructure, regulated systems, and high-reliability software will continue demanding traditional engineering discipline. The durable code that runs financial systems, healthcare platforms, aviation controls, and core infrastructure requires stability, auditability, and deep architectural planning that no amount of AI generation can shortcut.
The real insight is that software is not uniformly one thing. The industry is splitting. Some code will be generated, used, and discarded in hours. Other code will be maintained and hardened over decades. The skill is knowing which is which — and that judgment, ironically, is the one thing AI cannot yet automate.
Frequently Asked Questions
What is disposable software?
Disposable Software: Why AI Is Making Throwaway Code the New Normal covers the essential aspects of this topic, examining current trends, key players, and practical implications for professionals and organizations in 2026.
Why does disposable software matter?
This topic matters because it directly impacts how organizations plan their technology strategy, allocate resources, and position themselves in a rapidly evolving landscape. The article provides actionable analysis to help decision-makers navigate these changes.
How does the bifurcation: disposable vs. durable work?
The article examines this through the lens of the bifurcation: disposable vs. durable, providing detailed analysis of the mechanisms, trade-offs, and practical implications for stakeholders.
Sources & Further Reading
- Disposable Software — Andreessen Horowitz (Anish Acharya)
- Disposable Code Is Here to Stay, but Durable Code Is What Runs the World — Honeycomb (Charity Majors)
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR
- AI | 2025 Stack Overflow Developer Survey
- Vibe Coding Is Passe: Karpathy Has a New Name for the Future of Software — The New Stack
- The End of Software — Chris Paik
