ROMA (Recursive Open Meta-Agents) is a framework designed to simplify the creation of high-performance, hierarchical multi-agent systems. It targets engineers, researchers, and power users seeking to build complex agentic applications, offering benefits such as parallel problem-solving, transparent development, and extensibility.

How It Works

ROMA employs a recursive plan-execute loop. A task is first processed by an Atomizer to determine if it's directly executable or requires decomposition. If planning is needed, a Planner breaks the task into subtasks, which are then recursively fed back into the Atomizer. Executors handle atomic tasks, supporting LLMs, APIs, or other agents. Finally, an Aggregator collects and synthesizes results from subtasks into a coherent answer for the parent task. This approach enables parallel execution of independent subtasks and provides full transparency into the system's decision-making process.

Quick Start & Requirements

Installation is streamlined via a setup script (./setup.sh), with Docker setup recommended for isolation. Key dependencies include Python 3.12+ and optionally E2B API Key and AWS credentials for secure code execution sandboxing. The project provides links to documentation and a frontend accessible at localhost:3000 post-backend setup.

Highlighted Details

• Recursive Task Decomposition: Automatically breaks down complex tasks into parallelizable subtasks with dependency management.
• Agent Agnostic: Integrates with any LLM provider via LiteLLM and a unified agent.run() interface.
• Transparency: Offers stage tracing for detailed debugging and optimization.
• Tool Integration: Seamlessly connects external tools, including E2B sandboxes for secure code execution and S3 for data persistence.
• Benchmarks: Demonstrates performance on SEAL-0, FRAMES, and SimpleQA benchmarks.

Maintenance & Community

The project emphasizes community-driven development. Specific details regarding maintainers, sponsorships, or dedicated community channels (like Discord/Slack) are not detailed in the provided README.

Licensing & Compatibility

ROMA is released under the permissive MIT License, generally allowing for broad compatibility with commercial and closed-source projects.

Limitations & Caveats

The framework is currently in Beta. Pre-built agents are presented as demonstrations requiring minimal tuning rather than production-ready solutions. A technical blog and paper are mentioned but not yet available.