terezatizkova

I do agree but if we think SDK vs API, the APIs will be the ones that will more likely adapt on agents, not SDK. What if the API changes, do you need to reimplementing SDKs with wrappers?

just seen that someone posted my essay here, and just wanted to say thank you for the opinions. there is a lot of disagreement and lot of good debate, which motivates me to write an update on this.

Tereza

E2B | Engineers (Distributed Systems, AI, Product, Security, DevRel), Brand Designer | onsite in San Francisco or Prague | https://e2b.dev/careers

E2B is a fast growing Series A startup with 7-figure revenue. We've raised over $32M in total since our funding in 2023 and are supported by great investors like Insight Partners. Our customers are companies like Perplexity, Hugging Face, Manus, or Groq. We're building the next hyperscaler for AI agents.

Europe down

Thank you! yeah, good point, that’s how AI crawlers “see” the web

Can it also do a search through documentation? Cursor agent can do that.

We just launched a secure virtual desktop made specifically for LLM-powered computer use. The difference from Claude computer use is that:

It works with any LLM

It has high level of security, because each sandbox instance is isolated from the others, and it's isolated from user's os.

The E2B sandbox in general can be customized with any dependencies or pre-installed packages.

Some features:

- Python or JS/TS version

- Allows streaming of the desktop screen

- Supports mouse & keyboard control, opening files, screenshots, running bash and PyAutoGUI commands & more.

Here is the Desktop Sandbox repo: https://github.com/e2b-dev/desktop We also added an example of computer use, using the E2B Desktop Sandbox: https://github.com/e2b-dev/secure-computer-use

I will be happy to hear any feedback if you want to check it out.

Is it only for people with developer access?

Amazing! What is the logic behind doing this with firecrackers vs other solutions? From what I understood, it is especially helpful for the use-case of coding agents, as firecrackers are good fit for running untrusted code. But how exactly that works and are there any alternatives that you considered?

Is any of these SDKs good for real-time agents' debugging?

Love it!