AOrchestra
A research framework for automatically creating specialized sub-agents at runtime, modeling each as a four-part interface of instructions, context, tools and model rather than a fixed, hand-engineered role.
AOrchestra addresses a gap between two common sub-agent designs: treating sub-agents as identical, context-isolated copies with little specialization, or hand-engineering static predefined roles such as coder, searcher or writer that are hard to maintain across tasks. It instead models any sub-agent as a four-tuple of instructions, context, tools and model, concretized at runtime by a central orchestrator that decomposes the goal and synthesizes each sub-agent's tuple.
This decouples orchestration from execution, letting the system specialize sub-agents per subtask, explicitly control how much context each one shares to limit long-horizon degradation, and navigate cost-versus-performance trade-offs through the chosen context, tools and model, without relying on static roles or copying the full context into every sub-agent. Across the GAIA, SWE-Bench and Terminal-Bench benchmarks, the authors report a relative improvement of about 16 percent over the strongest baseline when paired with Gemini-3-Flash.
It is written in Python, and configuration includes example benchmark and model-provider settings to fill in with API keys.