Mental model
The agent has six first-class tool families: filesystem, execution, documents, web, browser, tables, plus MCP for external systems. It plans, executes, then reports — that is the difference from "AI chat."
Tasks it can actually do
- Sort 30 scattered docs into topic folders
- Scrape a webpage → save it → summarize → draft a newsletter entry
- Run your project tests, paste failing stacks back into the doc
- Use the Linear MCP to list "tickets assigned to me"
- Open Xiaohongshu via browser, scrape recent comments, classify them in a table
What it cannot do
It is not AGI. Hard judgment calls (should we email the client?) still need your review. Every disk write is shown as a diff first. That is a deliberate design, not a limitation.