hi friends! this year, i've been working on a general purpose, native terminal coding agent harness. i remember being absolutely struck the first time i ever used claude code and it could just do things. i'm also a big fan of crush and pi.

it's pretty early on. i still have yet to build all the higher-level behaviors that make something go from a harness to a proper agent. it turns out, making a coding agent is incredibly hard because you have to strap a development environment to an llm. in my hubris, i had thought "oh—it's just a terminal app that calls some api's!"

i'm trying to provide all the things that i think an agent should have out of the box: subagents, sandboxing, compaction, cross-platform support, a beautiful interface, etc. i would be absolutely delighted if you'd give it a go (as far as safety goes, the sandbox tends to be too restrictive rather than too lenient—i'm trying to figure it out as i go).

i made bestie because i don't like it when agents do things and don't tell you what they're doing. or when the harness itself won't provide visibility into things. so bestie is architected around being able to inspect anything that happens in an inspector pane that shows you the inputs/outputs of every tool call, background job update, and shell process.