armada
A CLI that turns a repository into a self-organizing AI engineering team of eight specialist agents, running a contract-driven, evidence-gated loop of dispatch, build, test and review until a pull request lands.
armada implements what it calls loop engineering: instead of one-shot prompting an AI coding agent, a developer writes a contract describing what to build and how to verify it, and a control loop dispatches, verifies, gates and iterates a fleet of agents against that contract until a reviewed pull request lands in the repository.
The loop enforces a maker/checker split, so an agent that wrote code is never the one that approves it; runs independent phases in parallel as background subagents with disjoint file scope, serializing only phases that depend on each other; requires evidence such as a passing test run, a screenshot or a file-and-line citation at every gate rather than a self-reported claim; and writes every state transition to disk, so a killed and reopened session resumes exactly where it left off. Feature work is isolated so parallel voyages do not collide on the same working tree.
It is a Node.js-based CLI, built around the OpenCode agent runtime.