ROMA (Recursive Open Meta-Agent) is Sentient’s open-source meta-agent framework for building high-performance multi-agent systems. Rather than hand-wiring agents for each task, ROMA recursively breaks a goal into subtasks, plans and executes them with specialised agents, and aggregates the results — the same meta-architecture applied at every level of the problem. Released in beta (v0.1), the project has more than 5,100 stars on GitHub and was actively developed in 2026.
Multi-agent systems often fail in the wiring: developers hard-code which agent does what, and the design breaks when tasks change shape. ROMA’s answer is a recursive meta-agent loop with defined roles — an Atomizer decides whether a task is simple enough to execute directly or must be decomposed, a Planner splits it into a dependency graph of subtasks, Executors carry them out, and an Aggregator synthesises the outputs — and this loop applies recursively to subtasks themselves. The result is a general scaffold for deep-research-style and complex workflow agents, where depth and parallelism emerge from the task instead of being designed in advance. ROMA comes from Sentient, the open AGI initiative, and underpins their open-source research-agent work; the framework is model-flexible and aimed at developers who want agent-of-agents behaviour without building the orchestration from scratch. As a beta release it is evolving quickly, with the repository documenting the architecture, configuration and example agents.
Free. ROMA is an open-source framework released by Sentient; you pay only for the model API usage of the providers you plug in.
Developers and researchers building complex multi-agent systems — deep research agents, workflow automation and similar — who want a recursive orchestration scaffold instead of hand-wired agent graphs.