Tag
This article introduces Onyx, a VM for programmable agent orchestration, turning agent orchestration into software engineering by providing a runtime with clear semantics and execution model. It discusses the design decisions and limitations of current agent systems, and presents the PROGRAMS specification.
An open-source team found that by stripping the orchestrator of search permissions in a deep research system, forcing it to engage in high-level strategic thinking, Onyx surpassed Claude and ChatGPT on the DeepResearch Bench, becoming the strongest open-source deep researcher.
The Onyx open-source deep research system achieves top ranking by stripping search access from its orchestrator agent, forcing it to decompose queries into focused research threads. Its three-phase pipeline and two-level architecture prevent information distortion and premature answering, outperforming proprietary solutions from OpenAI, Anthropic, and Google.