So let's say a big company was willing to spend the money to do this. After creating millions of pull requests who is going to review merge these millions of AI authored patches?

They would have rewritten all commercial software better - or if we are to believe the fantasy and hyperbole, you could patch your Windows 11 kernel live with some frontier model to fix every bug known and unknown.

First of all they cant - second of all they wontr, because their goal is money.

I attempted a very small version of this at work, on a small area of our codebase.

Firstly, how are we defining tech debt in the abstract to guide an agent?

Once it got going, probably half the PRs got thrown out, 25% seemed like preference changes, and 20% were of any value.

I'm not sure it scales

I would genuinely start with assessing which tech debt exists, this does not require LLM (e.g. use Enola). Then think, whether it's worth fixing.

Allowing an LLM to solve the tech debt will most likely lead to micro-optimisations, and in larger repos even very odd architectural patterns.

Cause it costs money?