DevBox
An autonomous engineering testing framework built on Deep Agents that adopts a fully agentic sandbox model. It automatically converts task descriptions into draft pull requests and incorporates credential isolation, state persistence, and structured handoff mechanisms.
Timeline (last 90 days)
DevBox is an autonomous engineering testing harness built on Deep Agents, with the goal of automatically turning tasks into draft PRs.
DevBox's sandbox contains only dummy secrets, and outbound traffic passes through a proxy layer called iron-proxy, which replaces the dummy secrets with real ones.
DevBox's sandbox itself is disposable, but the disk is not; LangGraph checkpoints are stored on persistent volumes partitioned by task.
If a Pod terminates unexpectedly, the newly started sandbox mounts the same persistent volume and resumes execution from the last LangGraph checkpoint.
In DevBox, agents do not pass files directly to each other; instead, the testing harness starts a worker for each repository, each worker returns a Pydantic-validated result object, and the harness validates it before passing it to the next stage.
Before planning, DevBox agents pull the latest code from all repositories to ensure they do not work from outdated code.
DevBox's original author raised the sandbox creation timeout from 30 seconds to 5 minutes.
DevBox's code has been open-sourced on GitHub under the MIT license.
When restoring from a LangGraph checkpoint, the most recent incomplete step may be replayed, so node logic must remain idempotent to avoid duplicate side effects (such as creating a branch twice or posting a comment twice).
All Facts (11)
DevBox是一款面向开发者的云开发平台，将应用的创建、开发、调试、部署和发布全流程简化
85%UnverifiedDevBox基于Sealos云操作系统，Sealos是以Kubernetes为内核构建的云操作系统，由国内团队开发并完全开源
85%UnverifiedDevBox's sandbox contains only dummy secrets, and outbound traffic passes through a proxy layer called iron-proxy, which replaces the dummy secrets with real ones.
50%UnverifiedDevBox's sandbox itself is disposable, but the disk is not; LangGraph checkpoints are stored on persistent volumes partitioned by task.
50%UnverifiedIf a Pod terminates unexpectedly, the newly started sandbox mounts the same persistent volume and resumes execution from the last LangGraph checkpoint.
50%UnverifiedIn DevBox, agents do not pass files directly to each other; instead, the testing harness starts a worker for each repository, each worker returns a Pydantic-validated result object, and the harness validates it before passing it to the next stage.
50%UnverifiedBefore planning, DevBox agents pull the latest code from all repositories to ensure they do not work from outdated code.
50%UnverifiedDevBox's original author raised the sandbox creation timeout from 30 seconds to 5 minutes.
50%UnverifiedDevBox's code has been open-sourced on GitHub under the MIT license.
50%UnverifiedWhen restoring from a LangGraph checkpoint, the most recent incomplete step may be replayed, so node logic must remain idempotent to avoid duplicate side effects (such as creating a branch twice or posting a comment twice).
50%UnverifiedDevBox is an autonomous engineering testing harness built on Deep Agents, with the goal of automatically turning tasks into draft PRs.
50%