Apr 24, 2025 · 1h 6m · latent-space
Why Every Agent needs Open Source Cloud Sandboxes
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
E2B co-founder Vasek Mlejnsky joins hosts Alessio Fanelli and Swix to discuss the architecture, scaling challenges, and strategic vision behind building open-source cloud sandboxes designed specifically for autonomous AI agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 31% of the talking time here. How this is scored →
The hosts as informed peer 5.3 Guest teaching 3.7 Guest disagreement 1.7 The hosts pushing back 1.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 40:11 Vasek calls out the SF bubble on LangChain's relevance
Vasek forcefully rejects the popular local Silicon Valley narrative that LangChain is obsolete, pointing out its massive 20 million monthly download metric.
Hardest push from the hosts ▶ 50:51 Swyx pushes back against bespoke web interfaces for agentsSwyx firmly rejects the idea that humans must redesign separate websites (llm.txt) for agents, arguing agents should instead learn to navigate human interfaces directly.
Biggest teaching moment ▶ 23:25 Vasek educates on untrusted LLM cluster lockoutsVasek educates the hosts on how autonomous LLMs break conventional shared cloud infrastructure by describing a real incident where an LLM locked engineers out of their own cluster.
The host holds their own ▶ 28:48 Swyx breaks down cloud pricing first principlesSwyx demonstrates deep domain mastery by detailing how every cloud infrastructure company must price compute, storage, networking, and control planes to survive.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Origins of E2B: From DevBook to Agent Sandboxes | 4 | 2 | 1 | 0 | Hosts open warmly as early investors and Discord peers. Vasek explains the origins of E2B out of DevBook's interactive browser documentation and how an early GPT-3.5 demo got viral attention from Greg Brockman. | |
| Smol Developer Experiments and the Code Interpreter Pivot | 5 | 3 | 1 | 1 | Swyx shares his firsthand experience running Smol Developer and notes model degradation over time. Vasek clarifies why headless Jupyter sandboxes were needed to maintain state for iterative LLM code execution. | |
| Scaling Hypergrowth: From 40K to 15 Million Sandboxes | 5 | 4 | 2 | 1 | Alessio references E2B's explosive metric growth from 40k to 15 million monthly sandboxes. When Swyx assumes deep research agents just use search, Vasek gently corrects him, demonstrating that tools like Manus use sandboxes as full horizontal runtimes. | |
| Positioning in the LLM OS and AI Engineer Landscape | 5 | 4 | 2 | 2 | Swyx asks about LLM OS positioning and challenges why standard cloud providers like Railway cannot fill this gap. Vasek explains how positioning E2B as a generic cloud computer failed until they targeted specific AI engineer workflows. | |
| AI Sandbox Architecture and Multi-Language Runtime Security | 4 | 5 | 1 | 1 | Vasek explains the core technical difference of untrusted LLM workloads, recounting how Hugging Face locked themselves out of their own cluster when an LLM altered permissions. Alessio notes the advantage of polyglot runtime composability across multiple languages. | |
| Technical Specifications and Usage-Based Billing Realities | 7 | 4 | 1 | 2 | Vasek describes unexpected customer workloads generating petabytes of data on free tiers. Swyx steps in with deep cloud infra expertise, outlining the first principles of pricing compute, storage, networking, and control planes from his Netlify days. | |
| Sandbox Persistence, Checkpointing, and Tree Search Frameworks | 6 | 4 | 3 | 3 | Vasek challenges the SF bubble narrative that LangChain is obsolete by citing its 20M monthly downloads. Swyx counters that the underlying chat completion paradigm is dying in favor of multimodal and real-time streaming architectures. | |
| Navigating MCP Protocols and Web Interfaces for AI | 6 | 4 | 3 | 4 | Vasek expresses healthy skepticism toward over-hyped comparisons between MCP and fundamental internet protocols like email. Alessio and Swyx bring data on Cloudflare crawl-to-visitor ratios and push back against creating fragmented secondary websites for agents. | |
| RL Training, Model Evals, and the AWS for Agents Vision | 6 | 3 | 1 | 1 | Discussion centers on reinforcement fine-tuning with Open R1 and evaluation benchmarks. Vasek praises Alessio for personally writing integration code for Berkeley's LMSYS Chatbot Arena, and outlines E2B's vision of becoming the AWS for AI agents. | |
| Relocating to San Francisco and Expanding Engineering in Prague | 5 | 4 | 2 | 1 | Swyx asks about moving from Prague to SF. Vasek debunks the myth of the 'Collison installation' by revealing Patrick Collison only did it three times, while explaining that customer feedback loops in SF were indispensable during the exploratory phase before rehiring in Europe. |
Statements from this episode (25)
Swix: Claude 3 degraded in capability a month after launch
“I used the same project to do this, to try to repeat the demo that I made for myself a month afterwards, and it wasn't anywhere as smart. So cloud three got dumber, but it looks like I like made up the demo or something, but no, like literally I just reran the…”
Mlejnsky: Agent sandboxes need persistent state, not one-off execution
“The important part is that you don't need to explain the model, and the model doesn't need to care about how to keep the state of the program running. So it was, especially with our earlier models, I think the models are now smarter, but they kept producing, l…”
Mlejnsky: Jupyter notebooks fail to scale for LLM code execution
“Jupyter itself isn't the right environment to actually do it because like because all the technical problems that you will run into once you are start doing, once you start doing it on scale. So it's, it gets slower and slower.”
Mlejnsky: Agent sandboxes are general code runtimes, not just interpreters
“Yeah, I think it's a good idea to stop thinking about a sentence was just for code interpreting. And more about like a runtime, code runtime for the LLM, or the agent. The use case for the sandbox, it's a very horizontal in a sense that it can cover everything…”
Fanelli: AI infrastructure is lagging application demand for the first time
“I think there's this usual, like, VCEO, we should invest in, like, the tools, you know, the picks and shovels instead of the application layer. But I think it's, this is, like, the first time where, like, the infrastructure is lagging the applications.”
Mlejnsky: E2B aims to be the Kubernetes for AI agents
“I think a good analogy here is sort of, like, technologically, it's kind of, you want to be the Kubernetes of the world for the agent, but with much better DX and easier, easier to use.”
Mlejnsky remains bullish on JavaScript and web developers driving AI adoption
“I still remain very bullish web developers and JavaScript world, TypeScript world, even though we have, like, a ton of usage from Python but there's so many web developers, and it's easier and easier to use LLMs. I really think that you need to cater to these …”
Mlejnsky: E2B sees 250k JavaScript and ~500k Python SDK monthly downloads
“From number of downloads of our SDK per month, it's like, 250,000 JavaScript, close to around half a million Python.”
Mlejnsky: E2B can kill and spin up a new sandbox in 150ms
“If that happens with us, we just like kill the sandbox and get, get a new one. He it takes like, I don't know, a 150 milliseconds.”
Mlejnsky: LLMs will soon configure and provision their own cloud sandboxes
“And the goal where we think this is getting going is the LLM like decides what it wants to do and how it wants to have the sandbox configured. So it's basically starts controlling the infrastructure itself and creating sandboxes themselves.”
Mlejnsky: An E2B customer has generated a petabyte of data
“We have a customer that produced petabyte of data. I mean, it's not free to host petabyte of data and that's growing.”
Swix: Failing to price compute, storage, or networking invites infrastructure abuse
“Everything can be broken down into some combo. Of compute, storage, and networking. If you fail to price one of them, they will, you will get abused because.”
Swix: Netlify's transition to usage-based billing took a senior engineer a year
“We ended up putting one of our most senior engineers on it. And she took a year to ship the whole building projects. That was presumably a board level objective, which was like, Hey, let's change from this pricing plan to this pricing plan. How hard can that b…”
Mlejnsky: LangChain gets 20 million monthly downloads and remains popular
“Well, I have a, I have, I don't know if this is an unpopular opinion, but, like, people keep telling me, ah, Langchain isn't popular, but if you look at its stats, like, it has twenty million downloads per month. How can you have not popular framework when it …”
Swix: Chat completions are dying and early AI frameworks will age poorly
“Like, I don't think people realize, but, like, I'll just say it out here, like, chat completions is dying. So, like, any framework that was built in that era with, like, no conception of real-time, no conception of omnimodal or multimodal native things, they w…”
Mlejnsky: Smarter LLMs reduce the need for prompt management tools
“Good question to ask when thinking all about dev tools with LM. Is my dev tool more relevant as the LMs are getting smarter, and as people need less prompting? I'm, for example, like, really bad at prompting. But like, I can get more work done over the years b…”
Vasek Mlejnsky: Comparing MCP to email protocols is far-fetched
“I see a lot of people, like, comparing to email protocols and such, which seems a little bit far-stretched to me at this current moment.”
Swix: Every DevTools company needs an MCP strategy
“I do think that every DevTools company needs some kind of MCP strategy for better or worse.”
Fanelli: Anthropic scrapes 6,000 pages per referral, compared to OpenAI's 250
“Google would be a two to one crawl to referral ratio, so for every two pages, they will read, they will send you one visitor. He said OpenAI is 250 to one, so they'll read 250 of your pages and send you one person. And Anthropic was like 6000 to one. So they'l…”
Mlejnsky: Hugging Face uses E2B sandboxes during Open-R1 RL training
“The way HuggingFace, who built the OpenROne project is using us is during like the reinforcement learn, code gen reinforcement learning step where the ROne model the OpenROne model has a training step where they give it a Code problem, and the model needs to g…”
E2B plans to let LLM agents deploy and manage apps directly
“Eventually, like, we want the LLMs to Deploy these services, apps that they are building, and, ah, have them manage it, and developer is more like in the backseat, like, looking at things if everything is working correctly. If your swarm of agents is working c…”
Mlejnsky: DevTool Startups in Existing Categories Do Not Need to Be in SF
“I think it's, ah, you can definitely build a DevTool company from Europe. I think it's a lot about question of how easy you want it to be in the earlier days, especially if you're building like a, you know, kind of like red ocean versus blue ocean waters. If y…”
Mlejnsky: Patrick Collison Said Stripe Did the 'Collison Installation' Only Three Times
“I asked about, like, three years ago when I had a chance to ask a question from Patrick Collison, like, how many times you did the Collison installation? And he was like, three times, but then after that you probably don't want to do that because you want to a…”
Mlejnsky: Remote European Engineering Works Once the Product Roadmap Is Clear
“Once once you have like a clear idea of what your product looks like, Then you can find really good expert on certain part of your infrastructure on, on database, and things like that, and just, like, have, like, top talent get top talent for that. The reason …”