Editing Openai/692dc40e-3b10-8004-b982-7bea1050182a (section)

==== Magentic-UI is built around a multi-agent architecture, each specialized for different tasks, plus a central orchestrator. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/blog/magentic-ui-an-experimental-human-centered-web-agent/|publisher=microsoft.com|access-date=2025-12-02}}</ref> ====

Here are the main components:
* Orchestrator — the “manager” agent (powered by an LLM), which plans tasks, breaks them into steps, delegates sub-tasks, and coordinates with you. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/blog/magentic-ui-an-experimental-human-centered-web-agent/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* WebSurfer — an agent that can control a headless (or sandboxed) browser: navigate pages, click, type, scroll, etc. Useful for automation of web tasks (searching, data extraction, form submission). Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Coder — an agent that can write and execute code (Python or shell), inside a sandbox (Docker container), enabling data processing, automation scripts, or more complex logic. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* FileSurfer — for file operations: reading, converting (e.g. to Markdown), analyzing documents, etc. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/blog/magentic-ui-an-experimental-human-centered-web-agent/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Optionally, MCP agents — custom agents that plug into the system, giving flexibility to extend functionality (APIs, custom tools, domain-specific workflows). GitHub<ref>{{cite web|title=GitHub|url=https://github.com/microsoft/magentic-ui|publisher=github.com|access-date=2025-12-02}}</ref>

On the interaction side, Magentic-UI supports several human-in-the-loop mechanisms. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Co-Planning — you and the Orchestrator draft a plan together, step by step. You can add, edit or remove steps before execution. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Co-Tasking — during execution, you can pause the agent, give feedback, intervene, or even take control manually (e.g. in the browser). Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Action Guards — especially for risky operations (code execution, sensitive web actions), the agent will ask for approval before proceeding. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/blog/magentic-ui-an-experimental-human-centered-web-agent/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Plan Learning & Retrieval — Magentic-UI can store completed plans in a “plan gallery,” so you can reuse or adapt them in future tasks — enabling automation of repeatable workflows. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/publication/magentic-ui-report/|publisher=microsoft.com|access-date=2025-12-02}}</ref>
* Parallel/Multi-tasking — supports running several tasks simultaneously (or managing multiple workflows), which can help for efficiency. GitHub<ref>{{cite web|title=GitHub|url=https://github.com/microsoft/magentic-ui|publisher=github.com|access-date=2025-12-02}}</ref>
* Long-running Monitoring (“Sentinel Steps”) — you can set up recurring or monitoring tasks (e.g. watch websites, APIs, or repo-updates) and have the agent trigger actions when conditions are met. Microsoft<ref>{{cite web|title=Microsoft|url=https://www.microsoft.com/en-us/research/blog/tell-me-when-building-agents-that-can-wait-monitor-and-act/|publisher=microsoft.com|access-date=2025-12-02}}</ref>