AI for building software

I have spent the past 4 weeks exploring AI for building software. It has been fun and enlightening. The output is this blog. I will be the first to admit that while I know more than I did 4 weeks ago, mine is still only a surface-level understanding. However, I am now in 100% agreement with the broad strokes of where software development is headed, given that agents and tooling have matured to a usable point and will only improve from here.

1. Software will be machine-made, not handmade.

2. The cost in $/hour to produce software and the time in hours/feature will drop 10x-100x. Because we can have practically infinite agents, running in parallel, 24x7x365. Not humans who take PTO.

3. There will be no 'badly written' software. But, there will continue to be differentiation between good software and great software.

4. Junior roles (not employees) are toast. Employees will reskill and find new roles -- by choice or by force. This will feel like graduating into the 2008 financial crisis or 2020 covid. But probably worse.

5. Software teams of the future will be lean. They will use external vendors to manufacture the software they want. A good proxy for the size of the future "in-house" software team is the following: Look at functions that carry out their job primarily through vendors. There are a handful of "in-house" employees, and for every in-house employee, there are anywhere from 10–100x "vendor" employees. Examples: Apple's manufacturing function (90–99% of the employees are at vendors, not Apple) and P&G's marketing function (90% of the employees are at agencies).

6. So who will be these external vendors that manufacture software? Many are throwing their hats in the ring: cloud-provider incumbents like Microsoft (GitHub Copilot) and Google (Antigravity); foundation-model leaders (Anthropic with Claude Code, OpenAI with Codex); and newer dedicated vendors like Cursor. My sense is that there will be enough space for everyone, but differentiation will be hard to establish, and being low-cost will matter. Basic functionality and quality are, of course, assumed. You can get a standard white mug for $1–$1,000, and while one may be better than the other, they both work just fine as mugs.

7. Handmade software will exist as a luxury product, like mechanical wristwatches, vintage cars, or designer handbags. There are obviously better ways to tell the time, get from A to B, or carry your stuff—but the above markets exist as status symbols. Maybe we get luxury software too.

Next 4 weeks: The two big buckets of building a company are building the product or service and building distribution/go-to-market. I plan to explore the second bucket and whether and how AI is changing this part of building a company.