Code And Let Live

5 points by teymour


zetashift

State management is hard, and having a nice "agentic workflow" suffers from it as well. But am I reading this right, they are starting a firecracker thingy for agents to solve this? So not like EC2, but more like ECS?

My first impression is that that doesn't sound sustainable? But I'd also rather see this solved at a level more similar to the BEAM/WASM or other runtimes.

But when it comes to sandboxing agents I'm going to guess this is more than enough for plenty of people.

lenny

Wanted to try this out but it looks like you can only access it via their cli, which you can only get from some install.sh that gets a binary from their servers. There appears to be no web interface or similar other than for creating access tokens.

I tried to check their github and I couldn't find the source for the CLI to install it myself.

So it might be proprietary?

Vaelatern

Is this the resurgence of image based development such as Smalltalk style?

sohalt

What's going to be interesting is how to make this kind of setup secure. With readonly ephemeral instances it's harder for malware to persist. With millions of persistent machines running untrusted code generated by ai agents on behalf of non-technical users I'm afraid we will see a similar pattern like all the pwnd WordPress instances running on classic shared web hosting with outdated PHP versions. The best we can do about this is probably strict sandboxing and some kind of IDS monitoring the sandbox from the outside for unusual behavior.