egg

egg turns an idea - a GitHub issue, a Jira epic, a rough description - into a reviewed, mergeable pull request, with agent teams researching the codebase, refining requirements, planning the work, writing code and tests, and reviewing each other's output across runs that can span hours and multiple repositories. Humans retain every decision that matters: whether requirements are right, whether the plan is sound, how ambiguity resolves, and whether anything actually merges.

The project's stated problem is that agents cannot be trusted with real credentials or a merge button, are unreliable at judging their own work, and quietly lose track of the original request as context windows fill and get compacted. Its answer is to move guarantees out of the prompt and into infrastructure: untrusted agents work inside a zero-credential sandbox, peer review must cite evidence from real build or test artifacts rather than trust the agent's own narration, and feedback gates sit at the cheapest point to redirect the work.

The project is under active development; egg is already used to develop itself end to end, including multi-hour, multi-slice pipelines, but the authors warn to expect breaking changes.