Ray Deck presents Magentic-1, an open-source multi-agent AI system with a master planner that delegates tasks to specialized agents for web browsing, file handling, coding, and command execution. It matches state-of-the-art performance on three benchmarks: Gaia, AssistantBench, and WebArena.
Alright, buckle up because I am absolutely buzzing to tell you about Magentic-1. It's a high-performing open-source generalist multi-agent system designed to conquer those truly complex open-ended tasks. At its heart is our brilliant orchestrator, the master planner, delegating subtasks to specialized agents: web surfer for browsing, file surfer for files, coder for code, and computer terminal for execution. It's a dream team, and the results? Phenomenal! It achieved statistically competitive performance against state-of-the-art methods on 3 challenging benchmarks: Gaia, AssistantBench, and WebArena, proving its incredible generalist capabilities and modular design. We also introduced AutoGenBench, an open-source tool for rigorous evaluation, and openly discuss its limitations, error modes, and the risks of autonomous AI. This is a huge leap.
