"If agents can build software on demand, why would you buy software at all?"

That question used to sound theoretical. It doesn't anymore. The way software gets built, deployed, and maintained is changing so fast that the build vs. buy framework we published recently already needs an addendum. Not because the framework was wrong. It still holds. But the ground underneath it is shifting.

This article covers three emerging concepts that will reshape the build vs. buy decision over the next two to five years: throwaway software, the dark software factory, and the idea that custom software could become a managed service. If you're making software investment decisions today, these are the trends worth watching.

Throwaway software

The idea is simple: instead of building software that lasts, you build software that does what you need right now and then disappears.

An AI agent receives a request. It writes the code, runs it, delivers the output, and the code is discarded. No maintenance. No technical debt. No versioning. The software exists only for as long as it's useful.

This isn't hypothetical. It's already happening in specific use cases. In the wealthtech and fintech space, platforms are using this pattern today. A user asks for a custom chart comparing three data sets over 36 months. Rather than trying to generate a static visualization, the AI agent writes actual code (Python, TypeScript, whatever fits), fetches the data, renders the chart, and executes it. The user gets exactly what they asked for. The code served its purpose and doesn't need to stick around.

It's a small example, but the implications are large. If software can be generated on demand for a specific task, a huge category of one-off tooling, internal dashboards, data transformations, and report generators stops being something you build or buy. It becomes something you ask for.

Why this matters for build vs. buy: The "buy" side of the equation has always included a category of lightweight tools that companies purchase because building them isn't worth the effort. Scheduling scripts. Data formatters. Simple integrations. If throwaway software matures, that entire category could collapse. You wouldn't buy those tools. You wouldn't build them either. You'd just prompt for them.

The limitation is predictability. When you run the same throwaway code twice, you want the same result. That's what engineers call deterministic outcomes. Agents still have reliability concerns around hallucination and inconsistency, which is why throwaway software works best today for analysis and visualization tasks where a human reviews the output before acting on it. For anything mission-critical or repeatable, you still want persistent, tested software.

The dark software factory

The concept borrows from manufacturing. A "dark factory" is a production facility that runs autonomously, no humans on the floor, no lights needed. The dark software factory applies the same idea to code.

The premise: you start with a prompt or a specification. A team of AI agents takes that input, builds the software, tests it against defined criteria, deploys it, and delivers a finished product. Humans define what to build. Agents handle the how.

As of early 2026, this is no longer speculative. BCG Platinion published a detailed framework describing autonomous software delivery as an emerging reality. StrongDM, an infrastructure access company, has operated a dark factory internally since mid-2025 with three engineers and a strict rule: no human-written code and no human code review. Spotify reportedly has engineers who haven't written a line of code since December 2025, using an internal AI platform called Honk to trigger and merge autonomous code changes.

Dan Shapiro, CEO of Glowforge, developed a five-level framework for AI autonomy in software development. Level zero is fully manual. Level five is the dark factory, where specs go in and software comes out. Most organizations today sit between levels one and three. But the gap between leaders and laggards is widening fast.

Is there hype? Yes. Some firms claim to be fully at level five today. The reality is more nuanced. But the trajectory is clear, and the distance between "AI assists the developer" and "AI builds the software" is shrinking with each model generation.

Why this matters for build vs. buy: The cost and timeline assumptions behind the build option are changing. If a dark factory pipeline can produce tested, deployable software from a spec in days instead of months, the traditional tradeoff (buy is faster, build is slower but more customized) starts to break down. Building custom software becomes faster and potentially cheaper, which means the threshold for when building makes sense drops significantly.

Software as a managed service

This is the most speculative of the three concepts, but it may also be the most consequential.

Today, most non-technical companies engage a managed service provider (MSP) for their IT infrastructure. That's the firm that sets up your Wi-Fi, maintains your email, manages your basic network services. It's a well-established model.

Now extend that model to software. What if a provider, whether an MSP or a new category of firm entirely, could deliver custom software on demand and maintain it for you? Not software as a service in the traditional SaaS sense, where you're one of thousands of tenants using the same product. More like instantaneous, real-time, bespoke software built for your specific needs, with a user base of one.

This would satisfy the need for custom software while keeping the distraction of building it off your plate. For non-technical companies, software could become liquid: something you consume as needed rather than something you own and maintain.

We're not there yet. But the building blocks are falling into place. Throwaway software proves that code can be generated on demand. Dark factories prove that the full build-test-deploy cycle can be automated. Combine those two capabilities with a service layer, and you get something that looks a lot like software-as-a-managed-service.

Why this matters for build vs. buy: If this model materializes, the build vs. buy question evolves into a three-way decision: build, buy, or subscribe to a managed software service. For companies whose core business has nothing to do with technology, this could be the path that makes the most sense. Custom software without the overhead of custom development. Differentiation without distraction.

What this means right now

These concepts are at different stages of maturity. Throwaway software is real and happening in production today, though limited in scope. The dark software factory is operational at a handful of pioneering organizations and quickly moving toward broader adoption. Software-as-managed-service is the logical extension, but it's still early.

Here's what hasn't changed: the decision framework from our Build vs. Buy AI article still applies. If a mature SaaS tool solves your problem, buy it. If your competitive advantage depends on proprietary data or custom logic, build it. If you need the power of frontier AI models without building your own, use the hybrid approach.

What's changing is the cost and speed of the "build" option. As AI agents get better at writing, testing, and deploying code autonomously, the timeline and price tag for custom software will compress. Features that once took months will take weeks. Projects that once required a 10-person team may require three people and a well-written spec.

At Fraction, this is already shaping how we work. We build the application layer on top of existing AI infrastructure. As the tooling for autonomous software delivery improves, our ability to ship faster and more cost-effectively improves with it. The pricing model stays the same: $149 per story point, scoped upfront, with a structured breakdown before you commit. What changes is how much more value fits inside each sprint.

The companies that will benefit most from these shifts are the ones paying attention now. Not to the hype. To the mechanics. Understanding what throwaway software can and can't do. Knowing when a dark factory pattern applies and when it doesn't. Recognizing that the build vs. buy equation is becoming more fluid, not less.

The goal isn't to predict the future perfectly. It's to make sure you're not making a five-year software investment based on assumptions that won't survive the next 18 months.

Related reading: Build vs. Buy AI: When to Build Custom and When to Use Off-the-Shelf, AI Strategy for Non-Technical Founders, What Is Agentic AI?

Sources

BCG Platinion, "The Dark Software Factory," March 26, 2026. Describes autonomous software delivery frameworks and early enterprise adoption, including Spotify and BCG Platinion's own legacy migration results.

Dan Shapiro, "The Five Levels: from Spicy Autocomplete to the Dark Factory," January 23, 2026. Five-level framework for AI autonomy in software development, from manual coding to fully autonomous delivery.

Simon Willison, "How StrongDM's AI team build serious software without even looking at the code," February 7, 2026. Documents StrongDM's three-engineer dark factory operating since mid-2025 with no human-written or human-reviewed code.

MIT Technology Review, "Generative Coding: 10 Breakthrough Technologies 2026," January 12, 2026. AI now writes up to 30% of code at Microsoft and over 25% at Google.

Anthropic, "2026 Agentic Coding Trends Report," January 2026. Examines how coding agents are reshaping the software development lifecycle.

‍