When is a browser agent still the better choice?
A browser agent is the better choice when no recipe exists yet, when the task is a one-off, or when the answer depends on what a page looks like. Wield recipes are faster and cheaper for the tasks they cover, and they run in your own browser. For everything else, agents still do the work that recipes do not cover. The two can be used together.
One-off tasks on unfamiliar sites
A recipe covers one task on one site. If no recipe exists, Wield cannot run that task yet. For a single question on a site that nobody has covered, a browser agent starts faster, because it needs no recipe.
Agents also cost more per task. A browser agent reads screenshots and clicks through each page. That takes minutes and uses model tokens for every step. If the task will come back, ask for a recipe with request_recipe from your agent, or record the task on a Pro or Team plan, so that later runs are fast.
Visual and multi-step exploration
Recipes make the same requests that the site's own pages make. They do not look at the page. If the answer depends on what a page looks like, a browser agent is the right tool. Examples include a layout check, a product photo that decides the choice, a site with no search form, or a path that changes at every step.
Exploration is also agent work. When the goal is "look around and tell me what is there", an agent can follow links and decide the next step. A recipe runs one defined task with set parameters.
Where recipes fail today
Recipes are not always the faster option. The benchmark shows where the Wield recipe passed fewer runs than it should, and where the browser agent passed fewer too.
| Site and task | Wield recipe passed | Browser agent passed |
|---|---|---|
| Zalando, fashion product search | 0 of 9 | 0 of 9, blocked in every run |
| ImmoScout24, rental flat search | 0 of 2 | 0 of 9, blocked in every run |
| mobile.de, used car search | 5 of 9 | 0 of 9, blocked in every run |
| Airbnb, holiday stay search | 8 of 9 | 4 of 4 |
Other limits apply as well:
- Browser support. Recipes run only in Chromium browsers: Chrome, Edge, Brave, Arc, Vivaldi, Opera and Chromium. Firefox and Safari are not supported.
- Open browser. Recipes run in a tab of the site, so your browser must be open. The window can sit in the background.
- Coverage. A recipe covers one task on one site. Anything else needs an agent or a new recipe request.
- Status. Some recipes are still pending a final check. The status on each site page shows where a recipe stands before you rely on it.
Using both
Wield is the layer that agents call. It is not an agent, so the two fit together. Let a recipe fetch structured data from a covered site, then let your agent reason over the result. Use a browser agent for the uncovered part, such as an unfamiliar page or a visual question. Results from Wield go to your AI app, which can use them in the same conversation as other work.
FAQ
Is a Wield recipe a browser agent?
No. Wield is not an agent, and no model runs when a recipe runs. It is the layer that agents call. The agent still decides what to ask, and a recipe answers that one request.
Which one costs less per task?
A Wield run costs $0 in model tokens. A browser agent costs its model's list price for each run. In our benchmark, the browser agent averaged $0.18 per task run. See how we benchmark for the method.
Which one is safer on a logged-in site?
Each has its own controls. A recipe runs only the requests it lists, on the origins it names, and every write stops for your Allow. A browser agent acts step by step on a page, so check what it does on a logged-in site before you let it act.
What if my browser is Firefox or Safari?
Recipes do not run in those browsers. Use a Chromium browser for recipes, or use a browser agent for the task.