Recursive agent teams that break problems down until they’re small enough

What it does ROMA is a Python framework for building hierarchical multi-agent systems. Every task enters an Atomizer that decides whether the job is atomic or needs splitting. Non-atomic tasks go to a Planner that breaks them into subtasks, which are fed back into the same pipeline recursively. Once the leaves are executed—by LLMs, APIs, or other agents—an Aggregator rolls the results upward into a coherent answer for the original parent task.

The interesting bit Recursion is a first-class citizen here, not an afterthought. The framework also tracks left-to-right dependencies between subtasks, so sequential work is supported alongside parallel branches. The authors claim benchmark wins for their search agent, though the README text is light on specific numbers.

Key highlights

• Recursive plan–execute–aggregate loop with dependency-aware scheduling
• Pluggable executors: any LLM, API, or agent implementing agent.execute()
• Ships with DSPy strategies (CoT, ReAct, CodeAct) out of the box
• Two deployment modes: lightweight file-based storage, or full Docker stack with PostgreSQL, MLflow, and FastAPI
• Defaults to OpenRouter (Claude Sonnet 4.5 + Gemini 2.5 Flash) but works with any provider

Caveats

• Explicitly beta software (v0.1); the authors note this is “just the beginning”
• Benchmark claims are mentioned but not quantified in the README text

Verdict Try it if you want a structured, recursive alternative to flat agent orchestration. Wait for a later release if you need hardened, production-grade tooling without rough edges.