What Magentic-One ships out of the box and when to use it versus building your own team
Magentic-One is a preset team (orchestrator plus four specialists) for generalist computer-use, useful as a baseline; build custom on AutoGen 0.4 when your workload shape diverges.
Imagine a moving company that arrives with a five-person crew already trained: one foreman, one driver, one packer, one loader, one heavy lifter. If your job is a typical apartment move, they just work. If you are moving a grand piano up six floors with no elevator, the standard crew is the wrong fit and you need to assemble a specialist team. Magentic-One is the standard crew for tasks that look like browsing the web, reading files, writing code, and running it. It is great when your job looks like that and frustrating when it does not. The framework underneath, AutoGen, is the staffing agency where you pick your own specialists.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Magentic-One is one of the most-cited reference systems in the 2024-2026 multi-agent literature, and it is also one of the most misunderstood. The misunderstanding follows a predictable shape: a team reads the paper, sees five agents working a hard task, and concludes that Magentic-One is the framework they should build on. It is not. AutoGen 0.4 is the framework. Magentic-One is a specific team that Microsoft Research assembled to demonstrate what AutoGen can do on generalist computer-use benchmarks like GAIA and WebArena.
This walkthrough separates the two, names exactly what Magentic-One ships, explains why the orchestrator's task ledger is the part worth copying, and gives concrete guidance on when to run Magentic-One versus when to drop down to AutoGen primitives and build your own team.
Mental model: Magentic-One is to AutoGen what a fully-furnished sample apartment is to a kit of building materials. The sample apartment is useful for seeing what is possible and for comparing your own design against; you do not actually live in it.
What Magentic-One ships, agent by agent
The orchestrator
The orchestrator is the manager. It does not have tools of its own; its only job is to think about the task, update its task ledger, pick the next specialist, and end the run when done. The ledger has four sections, rewritten every turn:
- Facts known (verified inputs and intermediate results).
- Facts to verify (open questions to resolve).
- Plan (sequence of upcoming specialist calls).
- Current step (what we are doing right now and why).
This ledger is the single most copyable idea in Magentic-One. Long-horizon multi-agent runs drift without it because each specialist only sees its handoff payload, not the global state.
The four specialists
- WebSurfer: a Playwright-driven browser agent. Can navigate, read pages, fill forms, click links.
- FileSurfer: reads and navigates local file system content (PDFs, Word, Excel, CSV).
- Coder: writes code (Python, shell) targeted at the next step of the plan.
- ComputerTerminal: a sandboxed shell that executes the Coder's code and returns output.
The Coder-Terminal pair is a generate then run loop; the SurfacePair (WebSurfer, FileSurfer) handles the read the world side. Together they cover the four most common verbs in computer-use tasks: read web, read files, write code, run code.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Microsoft Research's Magentic-One paper (2024) and the AutoGen 0.4 repo where the preset lives are the canonical references.
- Teams benchmarking on GAIA and WebArena routinely cite Magentic-One scores as the strong out of the box baseline to beat.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat does the orchestrator's task ledger contain and why is it the load-bearing piece?
Walk through the four ledger sections: facts known, facts to verify, plan, current step. Then explain that without it, a long-horizon multi-agent run drifts because each specialist only sees its immediate handoff payload and loses the global state.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating Magentic-One as a framework rather than a preset, and trying to bend it to a workload it was not designed for instead of dropping down to AutoGen 0.4 primitives.
60 second bullets to scan on the way to the call.
The five agents Magentic-One ships
What the orchestrator's task ledger contains
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.