Image: RudderStack
Why it mattersA shared coding agent that lives in your cloud and holds no real secrets can be reviewed like any teammate. The merge rate the owner reports gives a benchmark.
Teams have wanted a shared coding agent they can host themselves. The agent would pick up tickets and open draft pull requests. RudderStack released one on 28 October under the MIT license. The project is DevBox, at github.com/rudderlabs/devbox. The company has run it on its own engineers since April.
RudderStack says DevBox merged 734 pull requests across 60 of its own repositories in the last three months. Of the pull requests DevBox took to a decision, the company says 80 percent were merged. The rate rose each month: 72 percent in August, 79 percent in September, and 91 percent in October, counted to the 26th. The company says the gain came from a more capable agent and better briefs from its engineers.
How a task moves through it
A user assigns a Linear issue to DevBox, or mentions it on one. The control plane starts one sandbox per task through OpenSandbox. The agent pipeline then runs inside that sandbox. The pipeline has fixed stages. One agent reads the repository and its conventions. The next picks which repositories to change and writes a plan. Another agent argues with that plan. Implementation runs one agent per repository. Then up to six reviewers go over the diff at once. Each stage runs on its own model and reasoning effort. Review can run on a stronger model, while early stages run cheaper.
A review comment, a failed CI run or a Linear reply starts another round on the same branch. The sandbox holds only placeholder tokens. All outbound calls pass through a proxy. The proxy adds the real credentials and refuses any host not on the allowlist. A leaked token in the sandbox is worthless to anyone who finds it.
What runs where
The control plane is a Bun service with Postgres. The agents are built on DeepAgents. Sandboxes come from OpenSandbox, and egress goes through iron-proxy. Everything installs from one Helm chart. A laptop setup runs sandboxes as Docker containers. A production setup runs each one as a Kata Containers virtual machine on dedicated nodes. DevBox works with OpenAI-compatible endpoints and with Anthropic's API, so a team brings its own model keys.
The conventions DevBox follows live in AGENTS.md and skills files in each repository. The company says DevBox can only follow conventions a team has written down. The pull request that relicensed DevBox under MIT, ahead of this release, was itself opened by DevBox, the company says.
Most coding agents today sit inside one developer's editor. DevBox is the shared teammate shape. It has a credential story a security team can reason about. The owner has posted a merge rate for one real organisation. A team adopting it has a benchmark to measure its own work against.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.