Magentic-One: Microsoft's generalist multi-agent system tops GAIA benchmark
In one sentence Microsoft Research publishes Magentic-One: a system with an Orchestrator plus 4 specialized agents (WebSurfer, FileSurfer, Coder, ComputerTerminal). First place on GAIA benchmark. Key insight: stateless specialized agents plus stateful orchestrator outperform a monolithic agent. Open source MIT.
Microsoft Research presents Magentic-One, a system where five agents work together as a team: one coordinates, one browses the web, one manages files, one writes code, one runs terminal commands.
The interesting part is how it is designed: each specialized agent is stateless — it sees only the current task and remembers nothing from past interactions. All the memory and work plan live in the orchestrator, which knows where it is in the process and decides who to call each time.
This design proves much more robust than a single agent that must do everything: each specialist is more reliable in its domain, and the orchestrator can reassign work if one fails.
The result: on GAIA, the benchmark that measures capabilities on real tasks (answering complex questions requiring research, reasoning, and actions), Magentic-One achieves the best results. Released as open source under the MIT license.
Companies
Microsoft
Tools
—
Tags
Sources