Why Every Agent needs Open Source Cloud Sandboxes
Latent Space: The AI Engineer Podcast
2025/04/24
Why Every Agent needs Open Source Cloud Sandboxes
Why Every Agent needs Open Source Cloud Sandboxes
Latent Space: The AI Engineer Podcast
2025/04/24
Vasek Mlejnsky from E2B discusses the evolution and applications of sandbox environments for AI agents. The conversation explores how E2B has grown to support millions of sandboxes weekly, catering to Fortune 500 companies and various AI workflows. Key topics include the integration of multi-modality and tool usage in LLMs, as well as the platform's role in data analysis, code execution, and reinforcement learning.
E2B started as DevBook, offering interactive developer guides, before transitioning into a sandbox platform for AI agents. Early experiments with GPT-3.5 demonstrated the potential of running code in secure environments, leading to significant growth. By late 2024, E2B supported complex tasks like reinforcement learning and deep research. Despite challenges in aligning infrastructure with model capabilities, E2B adapted by focusing on specific use-cases and engaging with LLMOS companies. Python dominates for code interpretation, while JavaScript prevails in application development. Pricing strategies evolved from tier-based models to usage-based billing, addressing scalability concerns. Future plans include enhancing agent frameworks, integrating GPUs for faster data processing, and expanding into app hosting. E2B envisions becoming an AWS-like platform for LLMs, supporting the full lifecycle of AI applications. Relocating to San Francisco has facilitated closer ties with the tech community, aiding in recruitment and strategic growth.
00:03
00:03
Vasek prefers his name to be pronounced as 'Vasek'.
00:38
00:38
DevBook offered interactive playgrounds for developers, including one for Prisma still in use today.
02:40
02:40
They aimed to automate work by giving an agent tools to run code in a DevBook sandbox.
05:19
05:19
E2B was initially created to provide AI agents with a code-running environment.
07:27
07:27
The model became less capable over time despite initial promise.
10:35
10:35
Sandbox runtime for LLMs or agents is akin to a human using a laptop.
13:52
13:52
From 40,000 sandboxes in March 2024 to 15 million by March 2025
15:11
15:11
Deployment environments compared to Kubernetes for agents.
20:13
20:13
Python has nearly twice as many SDK downloads compared to JavaScript.
21:51
21:51
A story from Hugging Face demonstrates the importance of sandbox isolation.
26:28
26:28
E2B's Ubuntu sandboxes include a free tier with 2 CPUs and half a gig of RAM.
32:13
32:13
Hiring senior engineers for billing projects is extremely challenging.
34:08
34:08
Current agents' unreliability makes it difficult to price tasks in advance.
36:26
36:26
Forking and checkpointing can parallelize problem-solving and solve the local state problem.
41:57
41:57
A framework boom is expected as AI technology stabilizes.
42:35
42:35
Agent frameworks need to adjust to real-time, multimodal interactions and streaming.
46:47
46:47
DevTools companies must adopt an MCP strategy starting with APIs.
51:39
51:39
In reality, the world is messy, and clear distinctions between old and new entities are rare.
53:00
53:00
LLMs are likened to domain-name trends in the mobile era.
55:03
55:03
E2B sandboxes are integral to HuggingFace's reinforcement learning process.
56:58
56:58
E2B works with LMArena for comparing AI models in app generation
58:12
58:12
The GPU market is tough but offers new use-cases like faster data analysis and LLM training.
1:00:45
1:00:45
Aiming to build an AWS-like platform for LLMs
1:01:15
1:01:15
Building a DevTool company from Europe is possible but harder for fast iteration.
1:05:49
1:05:49
E2B is hiring for key roles to leverage market opportunities
