Researchers have introduced Magentic-One, an open-source agentic system designed to solve complex tasks through a multi-agent architecture. The system features a lead Orchestrator agent that plans, tracks progress, and re-plans to recover from errors, directing specialized agents to perform actions such as web browsing, file navigation, and Python code execution.
- Magentic-One achieves statistically competitive performance on the GAIA, AssistantBench, and WebArena benchmarks compared to state-of-the-art systems.
- The modular design allows agents to be added or removed without additional prompt tuning or training.
- The authors also release AutoGenBench, a standalone tool for rigorous agentic evaluation with controls for repetition and isolation.
This work demonstrates progress toward generalist agentic systems by showing that high performance can be achieved without modifying core agent capabilities or collaboration methods.