The first job to give a browser agent is the one you'd give a new temp

A Monday at a small service company

This is a composite scene, not a real customer. It's built from the kinds of conversations small businesses are having now that tools for browser agents are on the market.

Dana runs the office for a ten-person landscaping company. Every Monday she does the same three things before nine. She copies the weekend's online bookings into the crew schedule spreadsheet. She sends each crew lead a message with their jobs for the day. And she goes through the shared inbox to chase quotes that customers haven't answered.

The owner has read about AI agents that can do this kind of work in a browser. One product launched on October 20: WebRun says its agents click and type inside the tools a business already uses, like Gmail, HubSpot, QuickBooks and Xero, on a schedule, without connecting APIs or writing code. You can describe the job in plain language, record yourself doing it once, or start from one of its templates. The owner wants to know which of Dana's Monday jobs to hand over first.

Picking the first job

The tempting answer is the quote follow-ups. They're the most tedious, and they bring in money. They're also the riskiest. Each one is a message to a customer, in the company's name, and the right wording depends on what was quoted, when, and whether the customer already called.

The better first job is the booking copy. It happens every day, it has an obvious right answer, and if something goes wrong, Dana will see it in the spreadsheet within the hour. WebRun's own launch demo uses almost exactly this example: an agent adds each morning's new bookings to a schedule sheet and sends the team the day's list.

That's the test worth applying to any first job. The question isn't whether the agent can do the job. It's whether you'd notice quickly if it did it wrong. Copying bookings passes. Emailing customers doesn't, at least not yet.

Deciding what it can touch

Browser agents differ from older automation in one important way. They use a website the way a person does, which means they can, in principle, go anywhere a person can. So the second decision is about limits.

WebRun describes rules it calls policies. They check each proposed click, keystroke and page visit before it happens, then allow it, block it or pause the run for approval. There's an allowlist that keeps an agent to the websites you name, and a stricter preset that also blocks shell commands and file uploads. The company says the check runs on the action itself, so a blocked action stays blocked even if text in an email or web page tries to talk the agent into it.

For Dana's booking job, the setup is short. The agent may visit the booking system and the schedule spreadsheet, and nothing else. It may read bookings and add rows. It may not delete rows. The daily message to crew leads is set to ask first, so Dana approves it on her phone before it goes out. WebRun sends those approval requests through Telegram or WhatsApp.

Whatever tool you use, write these limits down in plain language before you set them up. "Reads bookings, adds rows, never deletes, asks before messaging anyone" is a rule the owner, Dana and the agent's settings can all agree on.

The first two weeks

For the first two weeks, Dana still checks the schedule each morning. It takes five minutes instead of forty. WebRun says every run is recorded, so when something looks off she can see what the agent did, in order.

She's looking for the predictable problems. A booking with a typo in the address. A customer who booked twice. A weekend job that was cancelled by phone but still shows online. None of these are agent failures, exactly. They're the judgment calls Dana used to make without thinking about them. Each one she spots becomes a rule: flag duplicates instead of copying both, skip bookings marked cancelled, put anything unusual in a separate list for her to look at.

For longer jobs, WebRun gives agents a worklist, like a list of 40 leads, and tracks which items are done, which are next and which were skipped, with a reason. Skipped items with reasons are exactly what Dana needs to see. They show her where the job still needs a person.

By the end of the second week she checks the log, not the spreadsheet. That's when the job has really been handed over.

When to give it the harder job

The quote follow-ups can come next, but only after the booking job has run cleanly for a while. Even then, start with drafts that wait for approval, not messages that send. Read every draft for a week. If Dana is rewriting more than a few, the instructions need work before anything goes out on its own.

It also helps to know the costs before expanding. WebRun's published plans are $20, $100 and $200 a month, which it describes as about 2,000, 10,000 and 20,000 agent tasks, with extra usage billed from pay-as-you-go credits. The Starter plan comes with a seven-day free trial, and the Business plan adds an audit log. Other browser-agent tools price differently, so compare by the jobs you'll actually run, not by the plan names.

What doesn't change

None of this removes the need for someone who owns the job. The agent copies bookings, but Dana still owns the schedule. If a crew goes to the wrong address, it's her phone that rings.

That's why the first job should be the one you'd give a new temp. It should be clear, checkable and low-risk, with someone watching the first few days and a short list of things the temp must never do. If an agent handles that well, you've learned how to hand over the next one.

If you'd like help choosing which jobs to automate, setting the limits and approvals, and connecting it all to the systems you already run, that's what our business automation team does. Talk to us.

Related work: Sutton Place.