How many requests does a task take: Wield recipe vs browser agent
A Wield recipe sends 1 to 4 requests per task, because it replays the site's own data requests. A browser agent loads pages the way a person does, so its requests include images, scripts and trackers. Across 23 benchmark tasks, the agent's per-task median was 204 requests. The highest was 924, on bahn.de. The data is as of 7 Nov 2026 and is preliminary.
Requests per task
Wield's count is the recipe's own data requests, median over its runs. The agent's count is every browser request in a run, median per task.
| Site | Task | Wield requests per run | Agent requests per run (median) |
|---|---|---|---|
| airbnb.com | search-stays | 1 | 361 |
| amazon.com | search-products | 1 | 320 |
| autoscout24.de | search-cars | 1 | 405 |
| bahn.de | search-connections | 3 | 924 |
| check24.de | compare-electricity | 2 | 470 |
| chefkoch.de | search-recipes | 1 | 187 |
| doctolib.de | get-slots | 1 | 615 |
| doctolib.de | search-practitioners | 3 | 688.5 |
| eventim.de | search-events | 1 | 14 |
| flixbus.de | search-trips | 3 | 251.5 |
| idealo.de | price-lookup | 2 | 2 |
| immobilienscout24.de | search-rentals | 4 | 21 |
| jameda.de | get-slots | 1 | 148 |
| jameda.de | search-doctors | 1 | 152 |
| kleinanzeigen.de | search-listings | 2 | 351.5 |
| lieferando.de | search-restaurants | 3 | 5 |
| mobile.de | search-cars | 1 | 5 |
| opentable.com | search-restaurants | 1 | 17 |
| vinted.de | search-items | 1 | 451 |
| walmart.com | search-products | 1 | 559 |
| willhaben.at | search-marketplace | 1 | 204 |
| zalando.de | search-products | 0 | 60 |
| zillow.com | search-homes | 1 | 17 |
Why the agent sends so many
The agent's count includes every request the browser makes: the page itself, its scripts and images, analytics and ad trackers, and the site's own data calls. Its steps also need a model, with 5 to 23 steps per run.
Only part of the total goes to the site's data. On bahn.de, one agent run sent 924 requests. Of those, 37 went to the site's API and pages.
What to watch in the numbers
- Blocked runs look light. Eventim.de (14), idealo.de (2), lieferando.de (5), mobile.de (5) and zillow.com (17) show few requests because the agent stopped early. Do not read them as efficient.
- Zalando shows 0 for Wield. The tab error stopped the run before any request. The success rate page has the failure details.
- Counts are not load. These numbers show how many requests each approach makes. They do not measure how much load either one puts on a site.
Method and caveats
- Basis: 23 benchmark tasks on 21 sites. bahn.de has one agent run, and the Zalando HTML fallback is excluded.
- Browser agent: Browser Use 0.13.14 driving Claude Sonnet 5.5, in a fresh headless Chromium with an empty profile.
- Wield: the recipe runs in the user's everyday browser through the extension.
- One agent and one model were used, so these counts are preliminary.
FAQ
How many requests does a browser agent make?
The median per-task count across the 23 tasks is 204 requests. The highest task median is 924, on bahn.de, from a single run. Blocked runs show far fewer requests because they stopped early.
Why does Wield need so few requests?
A recipe replays only the requests that return the data, usually one to four. The agent also loads every asset on each page and takes 5 to 23 model-driven steps to reach the same answer.
Does Wield load the site less?
The counts show fewer requests per task. They do not measure load on the site, so this page makes no claim about it.