TLDR: Sentient AI has launched ROMA (Recursive Open Meta-Agent), an open-source meta-agent framework designed to empower developers in building sophisticated AI agents capable of hierarchical task execution. ROMA aims to address the challenges of long-horizon, multi-step problems in AI by providing a transparent, modular, and recursive architecture for orchestrating multiple agents and tools.
Sentient AI has officially released ROMA, or Recursive Open Meta-Agent, an innovative open-source meta-agent framework poised to revolutionize the development of high-performance multi-agent systems. Announced on October 12, 2025, ROMA is specifically engineered to tackle complex, multi-step problems that often challenge traditional AI systems, with a strong focus on advancing Artificial General Intelligence (AGI) capabilities.
At its core, ROMA introduces a hierarchical, recursive task tree structure. This architecture allows a parent node to decompose a complex overarching goal into smaller, more manageable subtasks. These subtasks are then passed down to child nodes with relevant context, and their solutions are aggregated as results flow back up the tree. This systematic decomposition and aggregation process is crucial for reliably handling medium to long-horizon tasks that demand numerous sequential steps.
The framework’s design emphasizes transparency and traceability. By utilizing structured Pydantic inputs and outputs, ROMA ensures that the flow of context is explicit and fully traceable. This transparency is a significant advantage for developers, enabling easier debugging, efficient prompt refinement, and seamless agent swapping. Unlike ‘black-box’ systems, ROMA allows builders to clearly observe how reasoning unfolds, facilitating rapid iteration in context engineering.
ROMA’s architecture is comprised of distinct node types: the Atomizer, Planner, Executor, and Aggregator. The Atomizer evaluates whether a task can be directly executed or requires further decomposition. If complex, the Planner breaks it into subtasks, considering dependencies for sequential or parallel execution. Executors then perform these atomic subtasks, leveraging appropriate tools or agents like search APIs or extraction models. Finally, Aggregators combine and synthesize the outputs of these subtasks to form a coherent overall solution.
One of ROMA’s key strengths is its modularity, allowing developers to integrate any agent, tool, or model at the node level, including specialized Large Language Model (LLM)-based agents and human-in-the-loop checkpoints. The tree-based structure inherently supports parallelization, enhancing both flexibility and performance for demanding problems.
To demonstrate its efficacy, Sentient AI developed ROMA Search, an internet search agent built on the framework. This implementation achieved impressive results on benchmarks like SEALQA, reaching 45.6% accuracy and outperforming prior state-of-the-art systems such as Kimi Researcher. ROMA also showed strong performance on other benchmark subsets, including FRAMES for multi-step reasoning and SimpleQA for factual information retrieval.
Also Read:
- Information Theory Framework Unlocks Secrets of True Collaboration in Multi-Agent AI Systems
- OpenLedger Forges a Decentralized Future for AI with Blockchain Innovation
As an open-source and extensible platform, ROMA is designed for community-driven development. This approach encourages developers worldwide to contribute new agents, integrate specialized tools, and customize the framework for diverse applications, fostering collective improvements and innovation across the AI landscape.
