Wield recipe vs browser agent: how much faster, task by task
On the 13 benchmark tasks where the browser agent passed at least once, a Wield recipe took a median of 791 ms per run. The browser agent took a median of 70.6 s on the same tasks. The data is as of 7 Nov 2026. It is preliminary, and the benchmark is still running.
Key finding
The slowest Wield task on these 13 is check24.de at 1.5 s. The fastest is jameda.de get-slots at 211 ms. Every agent median is above 39 s. Across the full benchmark summary, the median run took 908 ms with Wield and 77,577 ms with the browser agent. That is about 85 times faster at the median.
Speed table
Median run time per task, from the benchmark's raw logs. Agent times include model calls and page loads.
| Site | Task | Wield median | Agent median |
|---|---|---|---|
| willhaben.at | search-marketplace | 582 ms | 162.0 s |
| jameda.de | get-slots | 211 ms | 46.0 s |
| doctolib.de | get-slots | 345 ms | 70.6 s |
| doctolib.de | search-practitioners | 891 ms | 160.5 s |
| autoscout24.de | search-cars | 860 ms | 144.5 s |
| kleinanzeigen.de | search-listings | 759 ms | 125.3 s |
| vinted.de | search-items | 920 ms | 136.5 s |
| jameda.de | search-doctors | 447 ms | 51.2 s |
| chefkoch.de | search-recipes | 414 ms | 39.5 s |
| check24.de | compare-electricity | 1.5 s | 103.4 s |
| flixbus.de | search-trips | 791 ms | 41.1 s |
| airbnb.com | search-stays | 1.4 s | 70.2 s |
| amazon.com | search-products | 1.3 s | 45.2 s |
Why the gap
The browser agent runs a model. A run takes 5 to 23 steps, and each step waits on a model call and a page load. A Wield recipe makes 1 to 4 requests to the site's own data endpoints. No model runs, so no step waits on one.
The two sides also run in different browsers. Wield runs in the user's everyday browser, which is warm. The agent runs in a fresh headless browser with an empty profile. The comparison measures the whole setup, not only the requests.
What it does not show
- It leaves out blocked tasks. The 13 tasks are the ones where the agent passed at least once. On the other tasks the agent never finished, so no agent median exists.
- It does not show success. Wield's failures are on the success rate page, including zalando.de (0 of 9) and immobilienscout24.de (0 of 2).
- The agent medians come from small samples. doctolib.de get-slots and kleinanzeigen.de have 2 runs each.
Method and caveats
- Basis: 23 benchmark tasks on 21 sites. bahn.de is included. Zalando's HTML fallback is excluded because it reuses the main Zalando agent runs.
- Each task has 3 fixed queries, and each query runs 3 times per side.
- Browser agent: Browser Use 0.13.14 driving Claude Sonnet 5.5, stock settings, in a fresh headless Chromium with an empty profile.
- Wield: the recipe runs through the real extension in Vivaldi or Chromium 152.
- Both sides must pass the same machine check.
- One browser agent and one model were used. Read these numbers as preliminary.
FAQ
How much faster is a recipe than a browser agent?
On the 13 tasks where the agent passed at least once, the median Wield run took 791 ms and the median agent run took 70.6 s. On the full benchmark summary, the median run took 908 ms with Wield and 77,577 ms with the agent. About 85 times faster at the median.
Is the comparison fair?
Not like for like. Wield runs in the user's warm everyday browser. The agent runs in a fresh headless browser, and it runs a model while Wield does not. The sample is small, with 2 to 9 agent runs per task, and the benchmark is still running. Read it as one preliminary setup.
What is the fastest Wield task?
jameda.de get-slots, at a median of 211 ms. The slowest of the 13 is check24.de compare-electricity, at 1.5 s.