rudderlabs

DevBox works like a junior dev. It shipped third of our PRs across 66 repos in past 6 months. Open Sourced [MIT] Source Code : rudderlabs/devbox MIT Licensed. No telemetry. Self-host it on your cloud. DevBox is an autonomous engineering harness. After setting it up in your own cloud, you can let engineering team delegate their tasks to DevBox. A typical workflow looks like this: 1. The engineering team member assigns a task to DevBox via CLI or Linear ticket (contributions are welcome to add more native task source or build on top of its API) 2. Sandbox: DevBox spawns a sandbox for each repo and set it up for deep agents run 3. Orientation: DevBox discovers the relevant repos (from the list of repos you have given it access to) and the components the tasks touches 4. Planning: Runs parallel agents for planning - one per repo 5. Plan Review: An adversarial agent checks the draft plan ands asks planning agents for revision 6. Implementation: One agent per planned repo makes the needed changes 7. Review: An agent reviews the difff 8. Commit: Raises a draft PR for the engineering team member to review 9. The engineering team members provides feedback in the PR 10. DevBox works on the feedback and raises following commits in the same PR 11. Distillation: This is the most important part which makes the agent better with every PR. An agent turns the human feedback and its own notes from the run into a new knowledge commit in the repo's configured knowledge directory (.agents/knowledge/ by default). It makes the next DevBox PR better for everyone in the team, increasing the merge rates. Note: Agents never get access to your secrets like GitHub token or LLM API Key. A proxy adds these secrets on the request's way out of sandbox. My team has been dogfooding DevBox for past 6 months. Happy to share it with you all and learn your opinions. Us it and if found a missing feature, contribute to it.

I went full-agent-in-sandbox with Deep Agents for DevBox, a task to PR harness While reading the sandbox-as-tool vs full-agent-in-sandbox breakdown - reddit.com, thought of sharing some nuances for my choice. I went with the full agent in sandbox for DevBox (autonomous engineering harness on top of deep agents). It is a harness to turn task into draft PRs. And for each task, I run the deep agent inside a sandbox. - Credentials: The sandbox holds only dummy secrets. On the way out for the traffic to GitHub and the LLM APIs, it goes through iron-proxy, which swaps them with the real secrets. - State: The sandbox is disposable. The disk is not. LangGraph checkpoints live on a per-task volume. If the pod dies, a new sandbox mounts the same volume and resumes from the last checkpoint. - Handoffs: An agent never passes files to the other. The harness starts one worker per repo. Each returns a Pydantic result that the harness validates and hands off to the next stage. - Tradeoff: A cold start per task. Each task gets a fresh pod and a volume restored from a periodic snapshot of the repos, then fetches every repo before it plans, so that it never works on stale code. Restoring the volume was slow, so I raised the sandbox create timeout from 30s to 5 minutes. Check the source code yourself: rudderlabs/devbox (MIT) Have you run. sandbox as a tool with 3+ subagents? Did it hold up with the shared state? ------ Disclosure: used ai to design the architecture diagram.

open-sourced by @rudderstack: devbox, which it says merges ~35% of its internal prs at an 80% merge rate. what it learns lands as a commit in the pr, reviewed like code. memory as a diff you can reject. would you trust agent memory you can't diff? rudderlabs/devbox

Linear ticket to a draft PR. A harness. Self-hosted, MIT - DevBox Disclosure: I'm an engineer at RudderStack. We have been using it for the past 6 months in our own team and open-sourced it this week - rudderlabs/devbox You assign a Linear ticket (or use the CLI) to DevBox. It plans the change, writes the code, runs the tests, reviews its own diff and opens a draft PR. Comment on the PR and it pushes the fix to the same branch. - Runs on your infra: Docker Compose, or a Helm chart on Kubernetes - Each task gets its own sandbox with placeholder tokens only. A proxy swaps in the real keys on the way out - Bring your own OpenAI or Anthropic key. - GitHub PRs only for now. On our own repos since April: 1,500+ PRs across 66 repos, about 80% merged. AMA. rudderlabs/devbox

工单等人工，PR 评论改完还得重推？ RudderStack DevBox 接 Linear 工单和 PR 评论，一任务一沙箱，真密钥不进沙箱，约 35% 的 PR 由它合并。 适合把改动交给 agent 的团队，人工负责 review 和合并。 哪类工单，修 bug 还是小功能？ #Agent #开源工具 @grok rudderlabs/devbox

I built a harness that everybody on my engineering team ended up using every day over the past 6 months. After delegating 1500+ Linear issues to it with an 80% merge rate, I'm open sourcing it (MIT license). I'm an engineer at RudderStack and I started building DevBox in April, and over the next six months everybody on our engineering team ended up using it. We open-sourced it this week (MIT License). My main frustration was that there is always some engineering work worth doing but never getting prioritized. And it keeps accumulating, too many Linear tickets in the backlog. So I built a harness to which I can assign these linear ticket and it will get back with a draft PR I'd actually merge. It plans the change, finds out relevant org repos, writes the code, runs the tests, reviews its own diff, and opens the PR. It works on the code review comments or CI failures as well and it pushes the fix to the same branch. What it learns about a repo, goes back to the repo as a commit on the PR and gets reviewed like code. Some notes about the architecture and the journey (top of my mind, ask me anything I miss) 1. As I deployed it to our own cloud, other engineers started using it. And it led to a shared knowledge improving our coding harness output at a faster pace. 2. My first version let a coding agent (Codex CLI) drive everything but that approach led to unexpected outcomes many times. Now the pipeline is fixed stages: orientation, planning, plan review, implementation, self-review. No code gets written until the plan passes review. 3. I replaced Codex CLI with Deep Agents. Giving us more control over choosing the model and reasoning effort for each stage. 4. Separated the agents from secrets and the control plane. The control plane never calls an LLM. It handles webhooks, task state and sandboxes. The harness runs inside the sandbox. 5. The LLM never sees the secrets (Github tokens, LLM API Key, etc.), they are added via a proxy on its way out from sandbox to the real API. 6. It struggles with vague tickets and big unspecified features. It's best on small, well-defined tasks. Repo: rudderlabs/devbox Demo Video: youtube.com This week, I open sourced it, MIT license. Deploy it in your cloud with Docker/Kubernetes, bring your own models (OpenAI/Anthropic supported for now), no telemetry. Let me know your experience.

@obiabo_immanuel @linear @github Yes, contributions are welcome. Find the issues that you would like to contribute to github.com/rudderlabs/devbox/issues

Today we open sourced DevBox under MIT license. About 35% of our internal PRs are merged by DevBox now. You can run it on your own machine or run it in your cloud. Works natively with @linear and @github rudderlabs/devbox t.co

We are open-sourcing DevBox, the autonomous engineering harness our engineers have run every day since April. Delegate the issue, review the pull request. Runs in your cloud. MIT licensed. rudderlabs/devbox t.co