ROMA

Why we included this project

ROMA is for teams that need to split genuinely hard problems into subtasks without hand-building the orchestration layer themselves. The framework wraps a recursive plan and execute loop in named modules: an Atomizer that decides whether a task needs splitting, a Planner that maps out a dependency graph of subtasks, Executors, an Aggregator, and an optional Verifier. Because it sits on top of DSPy, you can swap in any LLM provider and use any of DSPy's prediction strategies, which makes parallel decomposition and merging results feel less like writing control flow by hand. It also scales cleanly, from a pip install for local experiments to a Dockerized deployment with PostgreSQL persistence, MLflow tracking, an S3-compatible store, and a REST API. That puts it somewhere between a bare orchestration library and a heavyweight agent platform.