Thoughts on “Evolving programming languages in the AI era”?

Most LSPs, including Expert and Dexter, have a database inside but, as far as I know, none of them expose it via a query language.

And a challenge with such database is that it is constantly needing to reconcile with the data on disk.
Reminds me how both SQLite3 and DuckDB are able to query over data stored elsewhere.

Generally I think what’s good for humans will be good for LLMs

Not necessarily. Sometimes what would be best for both humans and agents is actually too tedious to stick to for humans in practice. See Elm or even DHH’s Rails World 2026 keynote about moving to Rust.

A great of example of targeting LLM capabilities and agents is the MoonBit language. The team is doing a fantastic job building tooling and tweaking the language itself to optimize for agentic outcomes. And we’re seeing more and more of these ideas in the Elixir world already, because it does help.

On the topic of types: As a human I love having types right there in the code, documenting without the risk of being stale.

it looks like someone already built something similar months ago: elixir-ontologies/guides/users/getting-started.md at main · pcharbon70/elixir-ontologies · GitHub (query an RDF knowlege graph)

Oh hey, a José jump scare

EDIT within the timeframe before it even shows I made an edit: Looking at your activity I now realize you’ve been pretty active relative to the recent activity on this forum so, uh, my joke sucks

The bit on “stronger guarantees” gave me pause.

Stronger guarantees is why I think that Rust is going to be enormously important as we move into agentic coding. You can’t really trust an agent to write memory-safe C++, for example, but in Rust you don’t have that problem because it’s memory safe if it compiles (barring the use of Dark Arts…).

I guess that the capability to develop memory-safe code in older languages might arise with guardrails, but that’s going to require scaffolding and eat up context that could be used for other things. Which leads me to my next point: one big design consideration will be asking ourselves what we can offload to static analysis. Rust, in particular, allows us to offload memory safety almost entirely, so it has a huge advantage.

This is also why my RAG framework focuses so heavily on static analysis, especially the kinds of static analysis that catch agentic errors. For example, property testing, which enforces invariants thus defeating overfitting.