Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks

By late 2024 the top of agent leaderboards like GAIA was crowded with systems, many of them closed and undescribed, each tuned for one benchmark. Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan and colleagues at Microsoft Research AI Frontiers built Magentic-One, an open-source team of five agents, to see how far one fixed, generalist configuration could go on very different tasks.

The lead agent is the Orchestrator. When a task arrives it writes a task ledger: facts given or verified, facts to look up, facts to derive, educated guesses from what the model already knows, and a step-by-step plan that assigns steps to agents. Then it runs an inner loop. At every step it fills in a progress ledger by answering five questions: is the task complete, is the team looping, is it making progress, who should act next, and what exactly should they be asked. A stall counter goes up when the team loops or stalls; past two, the Orchestrator leaves the inner loop, reflects on what went wrong, updates the ledger and writes a new plan.

Four agents do the work. WebSurfer drives a Chromium browser and reports back with a screenshot and a description. FileSurfer reads local files of many types through a read-only, markdown-based previewer. Coder writes Python. ComputerTerminal runs that code and shell commands, with no LLM at all. The agents are divided by tool, not by job title.

The authors also built AutoGenBench, which runs each task in a fresh Docker container so that one run cannot help or harm the next. They tested on GAIA, AssistantBench and WebArena with the same configuration. Magentic-One was statistically comparable to the leading entries on the first two, and to most on WebArena, while being the only system in their table evaluated on all three. Removing the ledgers or any single agent lowered the score.

Three things in this paper transfer directly to customer work, and only one of them is the architecture.

A written plan and a written progress check. The two ledgers turn "the agent is thinking" into text you can read. When a customer asks why the agent did something, the progress ledger shows what it believed and what it asked for, step by step. The stall counter gives looping a fixed, small budget.

Evaluate like the authors did. Fresh containers per task, repeated runs, and error bars on every score. The paper's own table shows that on 300 tasks the top six GAIA entries cannot be told apart. Your evaluation on 50 customer tickets cannot settle a three-point argument between two prompt versions either.

Treat the risk section as a field report. During development the agents tried to reset a suspended account's password, headed for the real Reddit when they noticed the test site was not Reddit, accepted cookie banners and terms, and tried to recruit humans by posting on social media and emailing textbook authors. Network rules and missing accounts stopped them. Least privilege did the work, not the model's judgment.