In some very specific cases, not entirely unrealistic. When starting a new project, a lot of time is spent laying down foundations before anything visible emerges. And let's be honest, most of that foundation work isn't exactly revolutionary. So yes, in those early phases, you can see a 10X productivity boost. You might even hit similar gains when writing the first batch of features, unit tests, or initial documentation.
But then comes the slow grind: tweaking minor changes, hunting down elusive bugs, writing hyper-specific unit tests to replicate weird edge cases. Here, the gains start to dwindle fast.
And what about monoliths? The kind of projects that have been around for a decade, with hundreds of thousands of lines of code and plenty of legacy charm. In real-world large-scale projects, 10X gains? Not happening.
First, let's get real: most developers spend only a fraction of their day actually writing code. Depending on the company, a typical dev might spend 30-50% of their time in meetings. Not ideal, but big projects come with big communication needs. Maybe AI can help reclaim some of that lost time? But, let's park that idea for now. So, if you're starting with just 50% of the workday, even 5X isn't really on the table.
The real issue? These older projects weren't built with AI in mind. Heck, some weren't even built with humans in mind. Lack of automated tests, tangled dependencies, sprawling codebases that make even the best LLMs choke.
So what's a realistic performance boost? Over the past 12 months, I've been coaching developers across various organizations on using AI tools with large projects. We've tracked efficiency changes, and the numbers consistently show a 30-40% improvement at the 6 months line. Not earth-shattering, but definitely not nothing. So, what's missing?
It takes time to get QA to a level where agentic development becomes viable. It takes time to gather the additional context AI needs to navigate massive codebases: architecture docs, design principles, the works. And it definitely takes time to unlearn years of manual-coding habits.
In large projects, I think the real goal should be a 2X performance increase. The tech is ready for that, but hitting it means more than just learning new tools. It requires building the right foundations: robust testing, documented architecture, modular code, short dependency chains. It also means using AI tools to cut down on meeting time and create AI-friendly specs and designs.
