Magentic-UI

Before it touches anything, Magentic-UI writes a step-by-step plan for the task and hands it to you. You can edit the steps in the plan editor, add or drop them, or send feedback in plain text to steer the approach. Execution only starts once you approve, which means you can correct a bad strategy before it burns time. A saved plan can also be pulled up on the next task so you're not starting from a blank slate.

During a run, the agent narrates what it's about to do and what it just saw, like which button it's about to click or what it read on the page. At any point you can pause the task, type a correction or grab the browser and do the step yourself, then hand control back. This is the difference from a set-it-and-forget-it automation tool: you stay in the loop because you can see and steer each action as it happens.

The agent action guards stop the run for your approval before it takes actions that are hard to undo, such as submitting a payment or sending an email. You decide how often those checkpoints fire, from every risky action down to a lighter setting for trusted tasks. Nothing irreversible runs without your say-so, which is the core safety idea behind the whole project.

The browser session and the code the agent writes run inside a Docker sandbox, kept apart from the rest of your machine. That separation matters because the agent can reach live websites, and a separate environment limits what it can touch on your system if something goes wrong. You can also restrict which sites it's allowed to visit, so it stays inside the boundaries you set.

Magentic-UI splits the work across specialized agents rather than asking one model to do everything. An orchestrator plans and delegates, a WebSurfer controls the browser, a coder writes and runs scripts, and a file agent handles documents and folders. Because it's built on AutoGen and speaks Model Context Protocol, you can add your own tools and agents to the team. The project is also integrated with Azure AI Foundry models and agents.

The agent learns from runs that went well. It can save a finished plan into a gallery and retrieve it later, either automatically for a similar task or manually from the list. Over repeated work, like a weekly data pull or a standard form, this cuts the setup time and nudges the agent toward a workflow you already approved. Reuse is the payoff for steering it through the first pass.

You can run several task sessions at once, and a session indicator tells you which one needs your attention and which has finished. Because the agent relies on your input at checkpoints, running tasks in parallel keeps you busy with approvals instead of waiting on a single agent to move. It turns the human-in-the-loop model into something closer to supervising a small team.