Browser agent
A browser agent is an AI agent that operates a web browser. It reads the page, either as the DOM or as screenshots, and then clicks, types and navigates one step at a time. Because the model acts at every step, a browser agent works on a site without a custom integration, but each run is slower and costs model calls.
How browser agents work
The agent receives a task and a view of the current page. It picks an action, such as clicking a button or typing into a field, and the browser performs it. The new page becomes the next input. The loop repeats until the task finishes or the agent stops. Every step is a model call, so a multi-step task makes many calls.
Why it matters for AI agents
Browser agents reach sites that have no API and no integration. The trade-offs are speed, cost and repeatability. Each run makes new decisions, so results can vary. Sites can also block automated browsers.
Wield and browser agents
Wield is not a browser agent. It does not decide each step with a model, and it does not click through pages. A Wield recipe is a small public JSON file that describes the requests a site's own pages make. The recipe runs in your own browser, with no model in the loop.
We ran both approaches on the same tasks, with the same machine checks. As of 7 November 2026, these results are preliminary and the benchmark is still running:
| Wield recipe | Browser agent | |
|---|---|---|
| Passing runs, pooled | 183 of 208 | 63 of 147 |
| Success, equal weight per task | 85% | 52% |
| Median time per run | 908 ms | 77,577 ms |
| Model cost per run | $0 | $0.18 |
The browser agent was Browser Use 0.13.14 with Claude Sonnet 5.5, on stock settings, in a fresh headless Chromium. Recipes miss too. In the benchmark, the Airbnb recipe passed 8 of 9 runs, and the mobile.de recipe passed 5 of 9. Full method on the benchmark page.
Related terms
- AI agent: the broader category that browser agents belong to.
- Tool calling: how an agent requests actions.
- MCP server: a way to offer tools to an agent without a browser loop.
FAQ
Why do browser agents get blocked?
Some sites run bot protection that shows a check, such as a CAPTCHA or an "are you human" step, when traffic looks automated. In our runs as of 7 November 2026, bot protection blocked the automated cloud browser on bahn.de, Lieferando, Zalando, ImmoScout24, Idealo, Eventim, mobile.de, Walmart, Zillow and OpenTable. We say "in our runs" because results change by site and by run. Wield does not solve or skip these checks. When a site shows one, Wield asks you to open the site in your own tab.
Is Wield a browser agent?
No. Wield is the tool layer that agents call. A recipe runs fixed requests from a tab in your browser, and no model decides the steps while it runs.
Sources
- Browser Usegithub.com