Notes

The assumption I had was Armin Ronacher’s starting premise: that the enormous corpus of existing code would freeze today’s programming languages in place, because a new language would have to overcome it. He now argues the opposite — that agentic engineering opens room for new languages rather than closing it, and sketches what a language designed for agents instead of humans would optimise for.

Read it for the reasoning rather than the prediction. It is one of the few pieces written from how agents actually fail, which is a more useful starting point than another round of benchmark numbers.

Sean Goedecke’s argument: a model produces code far faster than it produces judgement, so the skill that decides the outcome is reviewing what it hands back — the same instinct that catches a bad design in a pull request, pointed at a much higher volume.

It names the failure mode these tools are best at hiding: the code compiles, the tests pass, and the structure is wrong. That is worth remembering when the headline number is how much code was generated.

Mitchell Hashimoto’s account of adopting AI tools is unusually concrete: six named steps, from dropping the chatbot through to always having an agent running, each with what it replaced and what it cost him.

The order is the useful part. Reproduce your own work before outsourcing it, and build the harness before trusting it at scale — two of the six steps are about engineering the setup rather than writing better prompts, which is roughly where the effort lands when you measure agent runs instead of reading announcements.

Simon Willison updated ttok, his command-line tool for counting tokens with OpenAI’s tiktoken library, after leaving it alone for about two years: a Click warning fixed, CI refreshed, and a new --list-models. It runs without installing anything — cat file.txt | uvx ttok.

Worth keeping within reach for the same reason a kitchen scale is: the count is what turns “this prompt is long” into a number you can act on, and guessing is how per-task cost estimates go wrong. Publishing after two years of silence is also a decent argument that a small tool does not have to be maintained to stay useful.

A plain gist ranking the most popular blogs on Hacker News in 2025. As of today it has 1,799 stars and 256 forks — for a list of links, which tells you how few of these exist that are compiled rather than pitched.

It is where the reading list on this site started, and the reason the feeds here now carry the whole post: the feed depth of every entry on it was measured, and the sites that survive social media are mostly the ones handing readers their text instead of a teaser.