Overview

E2B evolved from DevBook to become an infrastructure platform providing specialized sandbox environments for AI agents to execute code, transitioning from interactive developer documentation to supporting complex AI applications across multiple programming languages.

The company experienced explosive growth from 40,000 sandboxes in March 2024 to approximately 15 million in March 2025, positioning itself as "Kubernetes for agents" with a focus on developer experience and infrastructure agnosticism across different LLM models.

E2B's technical infrastructure offers dynamic, runtime-optimized environments with complete sandbox isolation, supporting workloads from 5 seconds to 5 hours with flexible pricing models specifically designed for AI computational needs.

The company faces unique economic and pricing challenges in the emerging AI agent space, balancing usage-based billing complexity with the unpredictable nature of AI agents while developing advanced features like sandbox checkpointing and forking for parallel problem-solving.

Originally based in the Czech Republic but now headquartered in San Francisco to be closer to users, E2B aims to become "the new AWS for LLMs" by focusing on both technical excellence and strategic user engagement through in-person "collision installation" approaches.

Content

DevBook to E2B: Origins and Evolution

Vacek introduces his background with DevBook, a startup focused on interactive developer documentation and sandboxes

- Created interactive documentation and playgrounds for developers - Built an interactive playground for Prisma as an example - Initial technology was essentially unscalable sandboxes

Transition to E2B in March 2023:

- Motivated by burnout and emergence of GPT-3.5 - Initially aimed to build an AI development automation tool - Created an agent that could pull GitHub repositories, write code, and deploy

Pivotal moment:

- Tweeted their project - Greg Brockman (OpenAI) retweeted, generating significant interest - Project received ~500,000 views in a few days

E2B Name Origin:

- Originally had no formal name ("the AI company") - "E2B" means "English to Bits" - Quickly shifted focus from AI agents to sandbox infrastructure

Core Hypothesis:

- AI code agents would need specialized environments to run code - Similar to how developers need laptops/workspaces

Early Development and Technical Challenges

Initially developed a small agent project for Chrome extensions

- Project unexpectedly attracted a different audience than originally intended - Allowed agents to plan work in markdown files, which became a common approach for research agents

Technical Challenges:

- Early models struggled with complex tasks - Observed inconsistent model performance across different time periods - Noticed models had difficulty maintaining program state during code execution

Code Interpreter Evolution:

- Focused on creating a sandbox environment for agents - Identified Python as an early strong language for AI model interactions - Developed a specialized runtime environment to support Large Language Models (LLMs)

Market Strategy:

- Initially used project as a test for agent sandbox environment - Pivoted to code interpretation, particularly for AI data analysis and visualization - Needed a REPL (Read-Eval-Print Loop) environment to maintain code context

Growth and Expansion

Began expanding beyond basic code interpretation around late 2024/early 2025

- Saw significant growth trajectory in 2025 - Gradually added capabilities like computer use and other agent functionalities

Sandbox/E2B Development and Growth:

- In late 2024/early 2025, Sandbox saw expanded use cases beyond code snippets - Notable applications include: * Reinforcement learning * Computer use scenarios * Data analysis * Deep research agents * General agent runtime environments

Growth Metrics:

- Scaled from 40,000 sandboxes in March 2024 to approximately 15 million in March 2025 - Significant growth acceleration after Sonnet 3.7 release

Strategic Positioning and Market Approach

Strategic Positioning:

- Positioning as infrastructure that enables LLM applications - Aiming to be "Kubernetes for agents" with better developer experience - Emphasizing infrastructure agnosticism across different LLM models - Focusing on generalized sandbox capabilities

Market Observations:

- Infrastructure development is currently lagging behind application layer innovations - Increasing ease of switching between AI models - Users increasingly want deployment flexibility (cloud/on-premise)

Key Philosophical Approach:

- Viewing Sandbox as a versatile "dev box" for agents - Enabling agents to perform complex tasks similar to human computer usage - Providing tools that help agents work more efficiently across various domains

Market Education and User Focus

Market Education and User Adoption:

- The team focused on educating developers about their AI platform's potential use cases - Recognized that developers often struggle to imagine possibilities with new technologies - Emphasized showing concrete, specific use cases to drive early user traction - Aimed to make complex AI infrastructure accessible, especially to web and product developers

User and Developer Focus:

- Target users are primarily AI engineers, web developers, and JavaScript/TypeScript developers - Not focused on infrastructure or ML engineers - Believe in making AI tools simple and approachable for developers who want to build products

Language and Usage Insights:

- SDK download metrics show Python usage (around 500,000/month) is about twice that of JavaScript (250,000/month) - Python dominates for data analysis and code interpretation - JavaScript leads for building web applications and frameworks

Technical Infrastructure and Capabilities

AI Cloud/Sandbox Differentiation:

- Offers a dynamic, runtime-optimized environment unlike traditional cloud services - Enables fast dependency installation and GitHub repository pulling - Supports workloads ranging from 5 seconds to 5 hours - Provides flexible pricing models adaptable to AI-specific computational needs

Sandbox Infrastructure and Security:

- Sandboxes are placed inside clusters with a unique security model - Code running in sandboxes is considered untrusted by default - Complete isolation between sandboxes is crucial to prevent cross-contamination - Security and observability are key challenges

Compute and Runtime Flexibility:

- Supports multiple programming languages (Python, Lua, R, C++, Fortran) - Can run any Linux-compatible code - Ability to switch runtimes mid-process without creating new sandboxes - Elastic infrastructure that can dynamically adjust resources (RAM, CPU)

Technical Specifications:

- Provides Ubuntu-based Linux environment - Free tier offers: * 2 CPUs * 0.5 GB RAM - Customizable options up to: * 64 GB RAM * 16 CPUs - Free storage - Optional desktop SDK with VNC support for human interaction

Pricing and Business Model Challenges

Technology Infrastructure Pricing Fundamentals:

- Core infrastructure can be broken down into compute, storage, networking, and a control/security layer - Proper pricing of each layer is critical to avoid being "abused" by users exploiting free resources

Usage-Based Billing Challenges:

- Implementing usage-based billing is complex and requires extensive engineering effort - Instrumentation of infrastructure is critical and technically challenging - Engineers often resist working on billing projects despite their strategic importance

Billing Provider Considerations:

- Current options discussed include Orb, Open Meter, Metronome, and Stripe Usage - Key evaluation criteria include: * Revenue cut/pricing structure * Integration complexity * Time investment required

Specific Billing Complexities:

- Must track and measure usage across multiple dimensions - Difficult decisions around usage limits (soft vs. hard) - Balancing user experience with financial sustainability - Need to handle edge cases like early-stage startups with minimal usage

AI Agent Economics and Advanced Features

Pricing and Economic Considerations for AI Agents:

- Pricing models for AI agents are evolving, focusing on: * Reselling tokens from base layer model providers * Comparing value delivered vs. cost of goods sold * Exploring potential for task-based pricing

Token and Billing Dynamics:

- Some providers offer tokens with: * High markup * Negative markup (subsidized by VC funding) * "Bring your own key" solutions (less popular)

Agent Challenges:

- Current agents are: * Unreliable * Unpredictable * Difficult to price ahead of time

Sandbox and Checkpoint Technology:

- Emerging features include: * Memory persistence * Ability to pause and resume sandboxes * Continuous session time currently limited to 30 days * Potential for parallel problem-solving through forking

Future Potential:

- Forking and checkpointing could enable: * Tree/graph-based problem solving * Preserving local agent state * Avoiding full session replay * Exploring multiple solution paths simultaneously

AI Development Frameworks and Tools

Framework and Tooling Landscape:

- The discussion focuses on AI development frameworks and toolkits - E2B team prefers a toolkit approach over a rigid framework due to rapid AI evolution - Frameworks are challenging to build given the fast-changing AI landscape

Notable Frameworks and Tools Mentioned:

- LangChain: Surprisingly popular with 20 million monthly downloads - Mastra: A promising TypeScript-first framework - Composio: A toolkit providing multiple tools - Browserbase/Stagehand: An elegant tool for website navigation with minimal APIs

Key Insights on AI Development:

- Chat completions are becoming less relevant - Future frameworks need to consider: * Real-time interactions * Multimodal capabilities * Streaming agent interactions - Developer tools must adapt as LLMs become more sophisticated - The goal is to build architectures that can leverage significant model improvements

Machine Collaboration Protocol (MCP) and Web Interactions

MCP (Machine Collaboration Protocol) Discussion:

- Participants are exploring the potential and current state of MCPs - There's uncertainty about the primary use cases and target users (developers vs. end-users) - Current MCPs are mostly running locally, with emerging interest in remote MCP servers

Key Insights and Perspectives:

- Some view MCP as overly complex, comparing it to email protocols might be premature - Potential strategy involves creating higher-order MCPs that abstract implementation details - Developers believe every dev tools company needs an MCP strategy

Emerging Approach for MCP Implementation:

- Recommended to start with an API-first approach rather than an SDK - Need to distinguish between MCP clients and MCP servers - Authentication and persistent state are critical considerations

AI and Web Crawling/Referral Dynamics:

- Discussed crawl-to-referral ratios for different AI companies: * Google: 2:1 ratio * OpenAI: 250:1 ratio * Anthropic: 6,000:1 ratio

LLM and Website Interaction Insights:

- Current LLM text approaches are often automated website conversions - Suggestion that websites should have distinct experiences for humans and AI agents - Critique of current approaches as not truly separating human and agent interfaces

E2B Use Cases and Future Roadmap

E2B Use Cases:

- Highlighted use cases: AI data analysis, data visualization, coding agents, generative UI, code generation evals, computer use - Computer use considered most experimental and exciting, but currently limited in platform support

E2B Sandbox Use Cases and Future Plans:

- Hugging Face is using E2B sandboxes for code generation reinforcement learning, running thousands of sandboxes per training step - Benefits include: * Cost-effective alternative to GPU clusters * Isolated and secure environment * Enables large-scale model training parallelization

Emerging Use Cases:

- Model evaluation (e.g., running Sweetbench) - Comparing AI models (working with LM Arena from Berkeley) - Upcoming startup and research program for universities and researchers

Future Roadmap:

- Potential GPU offerings * Not immediate priority, but seen as a future opportunity * Could enable faster data analysis and machine learning model training - Potential app hosting and deployment services * Vision of LLMs doing development work with minimal developer intervention * Natural progression from current code execution platform

Company Location and Growth Strategy

Company Location and Strategy:

- Originally based in Czech Republic, the company moved to San Francisco - Primary motivation was to be close to users and be part of the emerging AI hub - Believed being in SF would allow more frequent, in-person user interactions

Hiring and Talent Considerations:

- Initially wanted to be physically present to understand product direction - Now comfortable hiring remotely in Prague/Czech Republic - Recognizes excellent technical talent exists in Europe - Currently hiring for roles in: * Distributed systems engineers * Platform engineers * AI engineers * Account Manager * Customer Success Engineer

Startup Philosophy:

- Advocated for "collision installation" approach - meeting users in person and quickly implementing their feedback - Referenced Patrick Collison's strategy of doing manual, unscalable things early on - Believes in-person meetings are more effective than remote interactions, especially in B2B

Current Growth Status:

- The company is experiencing market momentum and significant potential for growth - Strategic approach involves "doubling down" and accelerating progress - Metaphorically described as "pouring gas on the fire" to expand their current momentum - Aims to build "the new AWS, but for LLMs" - a comprehensive platform for AI development and deployment

Why Every Agent needs Open Source Cloud Sandboxes