Interesting experiment, but I was hoping you would provide some information that I could use to intuit how "the shape" of the output is different. What compromises have to be made to write IR directly? You mention that languages usually come with stdlib and other libraries to use, that you would not have access to. Seems like you kind of breezed over what libmuffintop actually is.
I enjoyed the writing style! For me tho, this felt a bit too process oriented and not enough result oriented.
Thank you for the kind words. Agreed that writing is process oriented and I'll have to work on that. This experiment has been on the top of my mind for a long time and it feels good to finally share it. It's been a long process.
Analysing the shape of the output is an excellent idea, it would be interesting to see which abstractions surfaced, and which did not. There are a lot of programming paradigms in high-level language that I'm curious to see if they are there or not.
And I do gloss over libmuffintop, and I think it might be worthy of its own post. It is the thinnest cross platform abstraction layer I could come up with. It's implemented in LLVM IR directly. The main idea is to remove as many abstractions as possible between the prompt and the machine. I wanted to see if the abstractions are necessary for agents, it seems that they are not.
Interesting experiment, but I was hoping you would provide some information that I could use to intuit how "the shape" of the output is different. What compromises have to be made to write IR directly? You mention that languages usually come with stdlib and other libraries to use, that you would not have access to. Seems like you kind of breezed over what libmuffintop actually is.
I enjoyed the writing style! For me tho, this felt a bit too process oriented and not enough result oriented.
Also point your agent at your website to try to reduce the font loading flicker. I did that on my site and it worked well https://hank.bond/microblog/2026-08-20-fout-reduction-is-a-g...
Thank you for the kind words. Agreed that writing is process oriented and I'll have to work on that. This experiment has been on the top of my mind for a long time and it feels good to finally share it. It's been a long process.
Analysing the shape of the output is an excellent idea, it would be interesting to see which abstractions surfaced, and which did not. There are a lot of programming paradigms in high-level language that I'm curious to see if they are there or not.
And I do gloss over libmuffintop, and I think it might be worthy of its own post. It is the thinnest cross platform abstraction layer I could come up with. It's implemented in LLVM IR directly. The main idea is to remove as many abstractions as possible between the prompt and the machine. I wanted to see if the abstractions are necessary for agents, it seems that they are not.