Most web agents ask you to trust them fully or not at all: hand over a goal, then watch a black box click around and hope it doesn't book the wrong flight. This one inverts that bargain. The interesting bet here isn't smarter automation — it's that the right amount of human control, applied at the right moments, beats raw autonomy on real tasks.
What Sets It Apart
- Four interaction modes make oversight cheap instead of all-or-nothing: you co-plan by editing the agent's step list before it acts, co-task by pausing to drive the browser yourself and handing control back, set action guards that force approval before irreversible steps like closing tabs or submitting forms, and let it learn reusable plans you can recall from a gallery.
- It runs as a team, not one model: an Orchestrator delegates to a WebSurfer for browsing, a Coder that runs Python and shell in Docker, and a FileSurfer for reading and converting files.
- Both the browser and code execution live in sandboxed containers, so the agent's mistakes stay contained.
How It Performs
On the GAIA benchmark, full autonomy completed 30.3% of tasks. With a simulated user available to answer questions, the same system reached 51.9% — a 71% relative jump — while asking for help on only 10% of tasks. The takeaway is that occasional, well-timed human input buys most of the benefit of constant supervision.
Who It's For
Great fit if you're researching human-agent collaboration, or want a controllable assistant for consequential web work where a wrong click costs real money or data. Look elsewhere if you want a fire-and-forget agent for bulk scraping or unattended automation — the whole design assumes a human stays nearby, and it's an experimental prototype, not a production service.
