ROMA – a multi-agent framework for Sentient AGI, open-sourced

What is ROMA？

ROMA (Recursive Open Meta-Agent) is an open-source multi-agent system framework developed by the Sentient AGI team. It decomposes complex tasks into parallelizable subtasks through a recursive hierarchical structure, where parent nodes allocate tasks to child nodes, which then execute and return results for aggregation. ROMA supports multimodal input and output, and comes with built-in agents such as a general task solver, a deep research agent, and a financial analysis agent, making it applicable to a wide range of scenarios from research and analysis to financial decision-making. With a transparent execution process that facilitates debugging and optimization, ROMA has shown excellent performance across multiple benchmarks, positioning it as a powerful open-source tool for DeepResearch.

Key Features

• Recursive task decomposition: Automatically breaks down complex tasks into parallelizable subtasks, integrates results after step-by-step resolution.

• Multimodal support: Handles diverse data types including text, images, and code, adaptable to various application needs.

• Tool integration: Supports MCP protocol and API integration, enabling calls to external tools and models.

• Transparent debugging: Provides clear visibility into every execution step, simplifying debugging and optimization.

• Built-in specialized agents: Includes a general task solver, deep research agent, and financial analysis agent to meet diverse requirements.

Technical Principles of ROMA

• Recursive hierarchical structure: Uses a tree structure where parent nodes decompose tasks into subtasks, which are executed by child nodes and returned to the parent.

• Core components:

  • Atomizer: Determines whether a task is atomic; if not, triggers decomposition.

  • Planner: Breaks down complex tasks into subtasks and recursively assigns them.

  • Executor: Executes atomic tasks by invoking LLMs, APIs, or other agents.

  • Aggregator: Integrates subtask results and returns them to the parent node.

• Context flow management: Top-down task decomposition and bottom-up result aggregation ensure clear information flow.

• Modular design: Allows insertion of any agent, tool, or model at the node level, offering high extensibility.

Project Links

• Official site: https://blog.sentient.xyz/posts/recursive-open-meta-agent

• GitHub repository: https://github.com/sentient-agi/ROMA

Application Scenarios for ROMA

• Research & analysis: The deep research agent automatically decomposes complex academic or market research tasks, integrates multi-source information, and generates reports.

• Financial decision-making: The financial analysis agent monitors cryptocurrency markets in real time and integrates multiple data sources to produce investment analysis reports.

• Project planning: The general task solver decomposes project tasks, assigns them, and tracks progress to support efficient project management.

• Enterprise automation: Builds multi-agent workflows to automate internal business processes, enhancing operational efficiency.

• Educational tools: Students can use natural language to create research agents that automatically collect and integrate information into research reports.