Thread by @bernhardsson

There's a lot of buzz about code sandboxes. Which makes sense – coding agents are very useful! However I suspect building a business around it is quite hard. Some thoughts in 🧵

First of all, the code execution itself is definitely hard, but not crazy hard – products like Firecracker and gVisor solve the underlying hard technical problem of isolation.

So what's going to matter long term? Some thoughts:
1. Doing this at very large scale. This is a hard systems problem, and it's also necessary to make the economics work. You need to run at least 100k or ideally 1M sandboxes on avg to have healthy revenue.

2. The SDK itself will be a strong differentiator. Developer ergonomics matter. Make it easy to get started and build powerful things
3. The technical primitives. Things like network tunnels and storage.
4. Performance. Starting containers in 100ms or less is hard.

5. State snapshotting. This makes it possible to suspend/resume/clone sandboxes which imo will make coding agents way more powerful (what if you can branch and run a whole search tree)

6. Adjacent products. In particular LLM inference (to output code), but also training – especially doing cool things with reinforcement learning (rollouts). This makes the economics work and also creates an end-to-end product.

End of 🧵 – but I should also mention that we're working on all these things at @modal! Very excited about code sandboxes as a market, but it's still very early.