Sentient AI releases ROMA: an open source and AGI-focused meta-agent framework for building AI agents with hierarchical task execution capabilities

Sentient AI has been released ROMA (Recursive Open Meta-Agent)an open source meta-agent framework for building high-performance multi-agent systems. ROMA structures agent workflows as Hierarchical recursive task tree: The parent node decomposes a complex goal into subtaskpassing them to child nodes as contextand later total their solution when results are reflowed– make Context flow is transparent and fully traceable Cross-node conversion.

Architecture: Atomization→Plan→Execution→Aggregation

ROMA defines the smallest recursive control loop. First is a node atomization Request (atomic or non-atomic). If it is non-atomic, planner Break it into subtasks; otherwise, a Executor Run tasks via LLM, tools/APIs, and even nested agents. one aggregator The child output is then merged into the parent’s answer. This decision loop is repeated for each subtask, producing a dependency-aware tree that executes independent branches in parallel and enforces left-to-right ordering when a subtask depends on a previous sibling task.

Information trends top down When tasks are broken down and bottom up when the results are aggregated. Rome also allows it Manual checkpoint at any node (e.g. confirming plans or fact-checking critical jumps) and surface Stage tracking– Input/output per node – so developers can debug and refine hints, tools and routing strategies and understand each transition. This solves a common observability gap in agent frameworks.

Developer interface and stack

Provided by Rome setup.sh quick start Docker setup (recommended) or Local settingsadd the flag E2B sandbox integration (--e2b, --test-e2b). stack list Backend: Python 3.12+ with FastAPI/Flask, Frontend: React + TypeScript and real-time WebSocket, LLM support: any provider through LiteLLMand Code Execution: E2B Sandbox. Data path support Enterprise S3 installation and goofy fusepath injection checks and secure AWS credential handling, while maintaining leaf skill interchangeability while meta-architecture manages task graphs and dependencies.

During development, you can connect ROMA to closed or open LLMs, local models, deterministic tools, or other agents without touching the meta-layer; inputs/outputs are defined as Padang Tik For structured, auditable I/O during runtime and trace.

Why recursion is important?

ROMA structure as Hierarchical recursive task tree: parent node Break complex goals into subtasksthrough them as contextand later total solutions for kids when results are reflowed. This recursive decomposition limits context to the scope required for each node, thus inhibiting rapid spread while Stage level tracking (Using Structured Pydantic I/O) Make the process Transparent and fully traceableso the fault is diagnosable rather than a black box. Independent siblings can run in parallel Dependency edges impose ordering, turning model/cue/tool ​​selection into controlled, observable components in a plan-do-aggregate loop.

Benchmark: ROMA Search

To validate the architecture, Sentient built Rome searchan Internet search agent implemented on the ROMA scaffolding (without claiming to be a domain-specific “deep research” heuristic). exist SEALQA (Seal-0)— A subset designed to emphasize multi-source reasoning — ROMA searches reported on 45.6% Accuracy, exceeding that of researcher Kimi (36%) and Gemini 2.5 Pro (19.8%). “ROMA” also reported state-of-the-art framework (multi-step reasoning) and SimpleQA is close to SOTA (fact retrieval). As with all vendor published results, these results are to be considered directional until independently reproduced, but they demonstrate that the architecture is competitive in reasoning-heavy and fact-centric tasks.

For additional background on SEALQA, this benchmark targets search-enhanced inference, where web results can be conflicting or noisy. Seal-0 focuses on problems that challenge current systems, consistent with ROMA’s emphasis on robust decomposition and verification steps.

The right place for Rome?

ROMA positions itself as the backbone of open source meta-agency: it provides Hierarchical recursive task tree The parent node decomposes the goal into subtasks by context Go down to the child node (Agents/Tools) and then total When results flow back. Design emphasizes transparency pass Stage tracking and support People in the loop checkpointwhile its modular nodes allow builders to plug in any model, tool, or agent and leverage Parallelization For independent branches. This enables multi-step workloads – ranging from financial analysis Creative generation – Easier design with clear context flow and observable execution.

ROMA is not another “proxy wrapper”, but it looks like a strictly recursive scaffolding: Atomizer → Planner → Executor → Aggregatortracking is performed at each hop, in parallel when safe, and sequentially when needed. Early ROMA search results are promising and consistent with the goals of the framework, but the more important result is developer control—clear task graphs, typed interfaces, and transparent context flow—so teams can quickly iterate and validate each stage. With the Apache-2.0 license and the inclusion of FastAPI/React tooling, LiteLLM integration, and implementation of sandboxed execution paths, ROMA becomes a practical foundation for building long-term agent systems with measurable, inspectable behavior.

Check code and Technical details.. Please feel free to check out our GitHub page for tutorials, code, and notebooks. In addition, welcome to follow us twitter And don’t forget to join our 100k+ ML SubReddit and subscribe our newsletter. wait! Are you using Telegram? Now you can also join us via telegram.

Asif Razzaq is the CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of artificial intelligence for the benefit of society. His most recent endeavor is the launch of Marktechpost, an artificial intelligence media platform that stands out for its in-depth coverage of machine learning and deep learning news that is technically sound and easy to understand for a broad audience. The platform has more than 2 million monthly views, which shows that it is very popular among viewers.