---
type: "post"
id: "how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c"
slug: "how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c"
title: "How E2B Uses Its Own Sandboxes to Develop E2B with AI"
description: "E2B says its engineering team routes backlog work through Claude agents running in E2B sandboxes: an architect prepares a task package, an engineer agent builds it, AI reviewers inspect the pull request, and a person signs off."
retrieval_nugget: "E2B says its engineering team routes backlog work through Claude agents running in E2B sandboxes: an architect prepares a task package, an engineer agent builds it, AI reviewers inspect the pull request, and a person signs off. The workflow illustrates how isolation and explicit verification can shape agent-assisted development."
published_at: "2026-11-03"
updated_at: "2026-11-03"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["claude"]
entities: ["E2B","Claude","Linear","Tomas Srnka","Manus Sandbox"]
related_patterns: ["ai-native-orgs-move-to-review-and-orchestration"]
novelty_label: "new_fact"
what_changed: "E2B details its internal Claude-agent development workflow: isolated task sandboxes and branches, architect-authored work packages, AI review plus human sign-off, and a demo PR in about 2.5 minutes."
prior_refs: []
source_url: "https://e2b.dev/resources/how-e2b-uses-its-own-sandboxes-to-develop-e2b-with-ai"
source_urls: ["https://e2b.dev/resources/how-e2b-uses-its-own-sandboxes-to-develop-e2b-with-ai"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c/","markdown":"https://newruntime.com/posts/how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c.md","json":"https://newruntime.com/posts/how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c.json"}
markdown_url: "https://newruntime.com/posts/how-e2b-uses-its-own-sandboxes-to-develop-6c3c0c.md"
---

# How E2B Uses Its Own Sandboxes to Develop E2B with AI

## Retrieval answer

E2B says its engineering team routes backlog work through Claude agents running in E2B sandboxes: an architect prepares a task package, an engineer agent builds it, AI reviewers inspect the pull request, and a person signs off. The workflow illustrates how isolation and explicit verification can shape agent-assisted development.

## Summary
E2B says its engineering team routes backlog work through Claude agents running in E2B sandboxes: an architect prepares a task package, an engineer agent builds it, AI reviewers inspect the pull request, and a person signs off. The workflow illustrates how isolation and explicit verification can shape agent-assisted development.

## What happened
In a talk presented at #nobullshit.camp 2026 in Bratislava, E2B Head of Engineering Tomas Srnka described how the company uses its own sandboxes to develop E2B. The talk was published November 2, 2026, and the source identifies it as a 28-minute talk. E2B builds an open-source runtime and cloud sandbox service, with a bring-your-own-cloud option also mentioned in the talk chapters. The account focuses on how its engineering team hands backlog tasks to agents rather than on a new product release.

The process begins with an architect harness that knows the codebase. A Linear ticket does not go directly to the engineer agent as a link: the architect turns it into a Markdown work package. The package states the intended outcome, where to look in the code, how to start the development environment and how to test the result. This makes the task more than a prompt; it carries the operational context and verification criteria the agent needs to work.

An engineer agent then builds the change inside an E2B sandbox, on its own worktree and branch. The talk describes the rule as one task, one sandbox, one branch. Work runs neither on an engineer’s laptop nor on a shared box or in production. E2B says templates can be built once and paused; each sandbox resumes a fork of the template, so software such as Chrome or the application started at build time is already running. The talk discusses a 100-millisecond startup target, but the supplied material does not establish that every task or configuration meets it.

After implementation, AI reviewers check every pull request for quality and bugs. A person still checks fundamentals before merge, with particular care on critical paths. The demo follows a real bug from the backlog through a work package to a pushed pull request in about two and a half minutes, and includes a screenshot as evidence of the fix. That is a single demonstration, not a general throughput benchmark. The company also says bugs caught by reviewers are fed back into the harness so agents stop repeating them. The source names Claude agents but does not specify the model versions, service pricing, or a quantified comparison with human-only work.

## What the archive adds
Our [Sandbox SDK note](https://newruntime.com/posts/sandbox-sdk-open-source-typescript-api-to-run-3c9597/) covered a shared TypeScript API spanning E2B and other sandbox providers. That was about making sandbox infrastructure portable; this is a new, company-reported account of how E2B uses its own environment in an engineering workflow. The stories are related by sandbox tooling, not the same event.

## Trend signal
This case supports the pattern that AI-native teams organize work around review and orchestration, rather than treating output volume as the primary measure. E2B describes an architect packaging tasks, agents producing changes, reviewers checking PRs and a person owning the merge decision. It does not report accepted-outcome rates or review time, so it illustrates the operating model without validating its performance claims. See [AI-native organizations move toward review and orchestration](https://newruntime.com/patterns/ai-native-orgs-move-to-review-and-orchestration/).

## Connections to other events
A January 17 Telegram post described Manus Sandbox as an isolated cloud computer created for each task and managed by an agent. The E2B talk gives a different angle: rather than describing a sandbox product in general, it shows how a team combines per-task isolation with a prepared work package, branch-level work and review gates. The earlier post is an adjacent example of the broader move to give agents their own execution environments; it does not establish that the systems or workflows are equivalent.

## Why it matters
The reusable part of this case is not the two-and-a-half-minute demo. It is the sequence of controls around the agent: a human- or harness-authored task package, a bounded execution environment, a separate branch, automated review, and a human merge decision. Those elements give teams concrete places to specify what “done” means and where to catch failures.

For a team considering coding agents, isolation can reduce the consequences of mistakes, while instructions on how to run and test the project make work more legible and reproducible. Reviewers can provide another detection layer, but they do not remove the need for human ownership—especially on critical paths. Feeding reviewer findings back into the harness is also a plausible way to improve future task guidance, though the source provides no measured reduction in repeat bugs. Teams should evaluate the full loop, including review burden and rollback risk, rather than treating a fast PR demo as proof of higher engineering productivity.

## Limits and uncertainty
The source is E2B’s own account of its internal process and a live demonstration. It does not provide a sample size, success rate, defect rate, reviewer false-positive rate, or comparison against engineers doing the same tasks without agents. The stated two-and-a-half-minute duration applies to one demo ticket; it does not show how long preparation, review, rework or production validation took. The material also does not identify model versions, pricing, broad availability of this workflow, or whether every project can use the same template setup. The 100-millisecond target is discussed, not demonstrated as a universal measured result.

## Reader takeaway
E2B’s case is best read as a workflow design, not a productivity benchmark. Give agents a bounded task package, a separate place to execute, and explicit ways to test their work; then keep review and merge accountability visible. The demo is a prompt to measure the whole delivery loop, not just time to first pull request.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)