Autonomous tasks

For open-ended work, you can give an AI agent a task with a duration and let it run autonomously — for example “collect listings for two hours” or “keep this dashboard refreshed for 30 minutes”. Instead of asking you what to do at every step, the agent makes its own decisions and keeps working until the time is up. This guide explains how that mode behaves so you know what to expect.

It applies when you drive a profile through the agent skill or the agent chat API.

The duration is a contract

When the task includes a work duration, that duration is a commitment: once the work starts, the agent runs to the end of the window without pausing to ask for confirmation. It only stops early on a genuine hard stop (see below).

One short round of questions up front. Before starting, the agent may ask a single batched set of questions — only things it genuinely can’t infer (which profile, which task, whether to resume an interrupted run). Then it begins.
No pausing mid-run. Every choice point — which page next, which link, how long to wait, how many actions — is decided by the agent itself.
Decisions are randomized. Counts, order, timings and dwell are varied run to run, so no two runs look identical and no single run feels scripted.
Steady status updates. The agent streams short progress lines (current page, action taken, elapsed vs. remaining time) at a regular cadence, so you can see it’s working without interrupting it.

Staying within context

Long runs can fill the model’s context window. When it approaches roughly half full, the agent compacts the run so far into a short state snapshot — what it has done, where it is, what’s next — and continues from that snapshot rather than ending early.

Hard stops

The agent ends before the duration is up only when continuing is impossible or unsafe:

A verification wall it can’t pass (phone, QR or similar).
The profile’s license has expired.
The browser crashed or the profile stopped.
A captcha wall that blocks further progress.

On a hard stop the agent reports what happened and where it left off, so you can resolve it and resume.

Strict profile scope

The agent works only inside the profile you gave it. It will not switch to, launch, or stop a different profile on its own — even if another profile would seem to fit the task better. If the assigned profile doesn’t match the task (wrong region, not running), the agent stops and asks rather than silently changing profiles.