Возвращение фабрик программного обеспечения: как ИИ автоматизирует массовое создание кода

Why software factories are back - and how they work in the age of AI Follow ZDNET: Add us as a preferred source on Google. ZDNET's key takeaways - Software factories will prepare prototypes for wide distribution.  - Leading AI companies are creating the foundations. - The specter of daily releases and updates demands greater automation. Tune into the TV show Shark Tank, and the most recurring question you hear voiced by the venture capitalists to budding entrepreneurs is whether they have arrangements with a factory to cost-effectively mass-produce their pooper-scoopers, comfort socks, or scrub pads. The idea of a mass-production facility may make sense for software delivery as well. Imagine submitting your prototype -- such as a vibe-coded initial version of an app -- to a service that will mass-produce your solution in a repeatable, automated fashion for delivery to a wide audience. Such is the promise of the "software factory." Also: This free Windows 11 tool can rescue your PC when things go wrong - here's how There's nothing new about the concept -- it was all the buzz about two decades back as more software development became componentized and repeatable. Microsoft was talking about software factories back in 2008. The idea faded for a while, but now it's back, fueled by the rapid production of software code generated and managed by artificial intelligence. Until recently, even with automation and DevOps practices, "coding created the bottlenecks for a lot of organizations," Moritz Plassnig, CEO of CloudBees, told ZDNET. "You had to hire more engineers, but it was very hard and costly to hire them." Today's software factory is built on AI models Thanks to the coding capabilities of foundation AI models and associated agents, the concept of the software factory has been revived -- with the automation of the software development lifecycle into repeatable chunks receiving serious consideration. With agentic coding, "we're less constrained on the coding side," Plassnig said. It can also elevate the roles of IT professionals. "Developer craftsmanship is not going away, as devs and engineers shift from writing code to exercising judgment around what gets built and shipped." Today's software factory concept is built around the AI models now on the market -- and the concept is almost the same across all the models. "Over the past 18 months or so, many of the companies that are on the cutting edge of using agents to automate their software development life cycle have been building the same machine, and independently coming to the same conclusion on what this machine looks like," according to Jaymin West, a forward-deployed engineer and technology evangelist, in a recent YouTube video. Also: 'I'm not a programmer' anymore: Linus Torvalds on the only two tools he uses now The companies building out software factories include the luminaries of today's agentic economy: Anthropic, Cognition, Cursor, Factory, Google, Github, OpenAI, and Ramp. "Every single one of these companies has converged on the same shape," West pointed out. "It's really important at this point to realize that these are the companies on the cutting edge and there is a clear pattern forming on how they are structuring their entire engineering framework." (Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) The software factory serves as a way to quickly act on ideas, Plassnig illustrated. "Let's say somebody works in a customer support role, and they have an interesting idea based on all the feedback they hear from customers. Like a car factory, you have a prototype that's experimental. But then you have a factory that produces the cars at scale. Likewise, that person with the software idea can submit it to the factory and the first iteration of that gets created. Then the product, engineering and IT professionals take a look. Then the factory can take it again and make it." Also: 'AI is an enormous tailwind for software companies': 5 tips for adapting to the new 'SaS' model The software factory helps address the challenge of validating and maintaining hundreds of thousands of lines of code, along with the software releases, updates, and refreshes that now take place on a daily basis at many organizations. "We will get to the point where as humans [we] can't validate what's working and what's not really working anymore," Plassnig said. "You want to know if software is actually working, if there are no bugs, if there's not an error rate that's spiking. Or maybe it's working, but is the agent's memory usage higher than you want? Today, you have humans looking at all those alerts, but then if you have 100 times more changes, like every few minutes, it breaks." What does a software factory look like? A typical software factory, as supported by the foundation models, is made up of six components, West outlined: - Queue: "Work arrives as an issue, not a prompt." - Control plane: "A durable surface, not a laptop." - Sandbox: "One per task. Destroyed after." - Pull request: "The output unit. Retrieved by a human." - Event stream: "Watch every action. Kill it without destroying the run." - Durable memory: "The sandbox does every run. What is not written to a file does not survive." Of course, software factories, no matter how efficient, are not a panacea for good software practices. Also: AI failed to properly patch software flaws 74% of the time, 1Password's study warns "It's no secret now that code generation is incredibly easy," said West. "It's incredibly cheap. Verification has become the blocker for a lot of these companies. Making sure that agents write code is very easy. But making sure that that code is correct is still a major engineering problem. And it's important to discuss why building a software factory doesn't mean you're just excusing verification or degrading the quality of your codebase."