The mythical well-specified project

I’ve been thinking about AI and autonomous agents. There are many variants of how to get to autonomy but they all seem to be based on the same premise: the “well-defined project”. This is why there is so much emphasis in a well-written PRD with extensive reviews, sign-off, etc. Essentially Waterfall 2.0.

I’m beginning to think that the opposite is true: the real automation is achieved with loosely-defined projects. Now, there is a difference between loosely-defined and loosely-specified. Most projects are well-defined (the requirements are clear) but poorly-specified (the requirements aren’t well communicated). This has been the main problem in the IT industry for decades, and I don’t think AI can solve this. AI can help accelerate the development side of things, but the real problem is in the iterative process that comes from trying to figure out what the true requirement actually is. We often say “people don’t know what they want”, but actually people do know what they want, they just don’t know how to articulate it. AI may be able to get to what they really want faster thanks to the quicker turnaround, but it won’t remove the human from the equation.

The mythical well-specified project, where you can get a piece of paper and reliably create the product

Loosely-defined projects are those where the requirements are inherently broad and vague. It’s the “I need something sort of like this”. That’s where you can let loose a bunch of agents to do things because they don’t need to exactly match a requirement. It’s like saying “a have a new pilates studio and I need a website where people can sign up for classes”. You don’t have a well-defined problem, any decent website where people can book classes will do. And AI can give you that very easily.

This is one of the reasons Anthropic et al. can claim to be so fast in their development. First, because they’re trading quality for speed (“Move fast and break things”, “Get big fast”, typical start-up mindset). But also, because what they’re doing is so new that it’s not well-defined. Any functionality that sort of manages your context can help. Any feature that sort of produces working code will do. It doesn’t need to do exactly this or that, just something roughly like this or that. This is why you can produce functionalities without wasting time in reviewing because, really, there isn’t anything to review against.

How many loosely-defined projects are out there? I don’t know, but I have a feeling that there are going to be more and more: with the new economic balance of AI-generated code more people will trade specificity for speed. How about the truly well-defined projects? Those will benefit from AI too, but not as much, and they will remain comparatively slower and more expensive.