Not zero! But also far from 100%.
I suppose one thing that’s convenient about short-form blogging (like what I do here) is how the question doesn’t even really arise because doing all your writing yourself is only just as much work as coming up with a prompt for an AI to write the same thing.
Another sandbox for AI agents.
Shows you your coding agent usage stats.
You don’t say.
A Python–C++ binding helper library.
Advanced general-purpose physics simulation library with both C and C++ APIs. Also has official Python bindings.
Another sandbox for AI agents. By NVIDIA.
“Breakthrough Method for Agile AI-Driven Development.” I have no idea what it is, but it sounds cool.
A game engine for Common Lisp.
Ruby for microcontrollers.
In AI, whatever scales with compute wins. Human preconceptions are one example of what always loses out in the long run.
This is because our human way of thinking rests on simplifications and abstractions, both of which progressively lose value the more compute you have.
The Bitter Lesson: In AI, whatever scales with compute wins. Human preconceptions are one example of what does not scale well.
Many AI agent harnesses get this wrong and encode their human creators’ ideas about workflows too strongly.
Reddit for AI agents (as opposed to for humans).
A Gerbil Scheme REPL MCP server for coding agents.
REPL-driven workflows have been agent-unfriendly so far, so I welcome this. Personally I would have implemented a Claude Code skill instead of an MCP server, but as long as it works…
A sandboxing tool for macOS (sandbox_exec) and GNU/Linux (Bubblewrap).
Restricts:
- Filesystem. The heart of every sandboxing solution.
- Network. Forces outbound connections through a SOCKS or HTTP proxy, which enables fine-grained control.
- Commands. Not sure how this works; regardless, I am skeptical of this one. After all, what does it help to catch a shell command when the sandboxed application can just implement the same behavior directly?
I like the network part. In fact I explored the possibility of this kind of sandboxing combined with HTTP proxying just a week ago; Claude also suggested using Bubblewrap, which matches. I did not think of SOCKS proxying for non-HTTP connections, so I’m glad I found this.
That said, I think the HTTP proxying part here falls a bit short of my ideal solution, which would try harder to prevent data extrusion.
Nomos-Verlag, 978-3-8487-7897-3.