I was inspired to start a new project this week building an agential harness from scratch. This is partly a pedagogical exercise so that I can better understand the mechanics of tools like Claude Code, which is now an essential component of my coding workflow, prompted by reading about Anil Madhavapeddy’s own experiments with bootstrapping an agential harness in OCaml. But I also think building agential harnesses that use local models effectively (as opposed to models hosted by frontier labs or compatible APIs) might be a space in which genuine research/hacking needs to be done.
After a bit of jockeying with Kagi Assistant (my LLM research platform of choice), I settled on the name konrad; following the naming inspiration of Anthropic’s ‘claude’, but with reference to the first name of a different computing luminary known for his autodidactical elan. I’m developing the codebase on tangled:
The idea is first to build an implementation in Rust, as this is the language with which I’m most familiar, and which seems most appropriate to the task at hand given that it will need to interface with relatively low-level code to run local models as close to the metal as possible. If that goes well, then I have in mind to see how successively I might develop alternative implementations in either/both OxCaml and Lean, two languages about which I’d like to learn more, from a language-independent spec that I’ll try to build alongside the initial Rust impl. This experiment is spirited by the idea that, given how well LLMs can now produce code, there might be newfound gravity in a software’s specification and test suite independent of any canonical implementation.