Just-bash is an interesting take on agent bash execution and I know that switching to bash commands was quite useful based on their blog post - but its quite different than what I expected. From what I understand based on the repo, they implemented bash commands + virtual filesystem entirely in TypeScript - agents can explore safely in-memory without any sandbox. Everything is simulated, so there's zero risk to your real system (is it really bash then if its just TS simulation of few commands?). Also this probably brings some limitations compared to real linux container, right? I'm wondering how quickly agents can hit the limitations of the simulation though - what happens when they need use some real binaries? Letting agents play in real sandboxes feels much more powerful for actual code execution imo. I know that the project is API-compatible with Vercel’s Sandboxes, so you can swap to real VMs when needed, but does it actually switch to real execution, or still use the simulation inside the sandbox? Anyway, really cool project - makes me wanna try this with real sandboxes and see the difference.
Bash is all you need! Which is why I'm introducing my holiday project: just-bash just-bash is a pretty complete implementation of bash in TypeScript designed to be used as a bash tool by AI agents. Because it turns out agents love exploring data via shell scripts, even beyond