Toggle the 🛠 Agent switch and Lumen stops just answering and starts working a loop: it plans a small set of steps, runs a tool for each one, looks at what came back, and decides whether to keep going, change course, or report the result. That's the same shape as the ReAct pattern -- reason, act, observe, repeat -- and each step is shown in the plan panel as it runs, not hidden behind a spinner.
What it can actually reach
- Web search and page fetches, for research questions.
- Connectors -- GitHub, Google and other services you explicitly link.
- A sandboxed, headless-browser container for tasks that need a real page (clicking, filling forms) -- it has no route back to other Lumen services beyond the one call it's allowed to make.
- A locked-down code-execution sandbox for calculations or quick scripts: read-only filesystem outside its own scratch space, no Linux capabilities, and -- unlike the browser container -- no network route out at all.
- Your own local files, but only after you explicitly pair a device on Connect local files -- nothing local is ever touched without that opt-in step.
Every tool call and its result is surfaced in the observation log in the UI, so "agent did something" is never a black box -- you can see the exact step, whether it succeeded, and what it found.