Success rate by site: Wield recipes vs a Claude browser agent
Wield recipes passed 183 of 199 benchmark runs across 23 tasks, or 92%. A Claude Sonnet 5.5 browser agent passed 63 of 147, or 43%. Wield failed outright on zalando.de and immobilienscout24.de, and partly on mobile.de and airbnb.com. The data is as of 7 Nov 2026 and is preliminary.
Key finding
The pooled counts above cover 23 tasks, with each run weighted equally. bench.json's own summary counts 24 tasks, including Zalando's HTML fallback (208 Wield runs, 183 passing), and weights each task equally. On that basis Wield passed 85.2% and the agent 51.6%. Both bases are valid. This page uses the pooled counts unless it says otherwise.
Wield had every run passing on 19 of 23 tasks. The agent had every run passing on 11 of 23, and it had no passes on 10 tasks.
Per-task pass table
Passing runs over total runs, per task. Each side uses the same machine check.
| Site | Wield | Browser agent |
|---|---|---|
| airbnb.com | 8 of 9 | 4 of 4 |
| amazon.com | 9 of 9 | 4 of 4 |
| autoscout24.de | 9 of 9 | 5 of 5 |
| bahn.de | 9 of 9 | 0 of 1 |
| check24.de | 9 of 9 | 5 of 5 |
| chefkoch.de | 9 of 9 | 8 of 8 |
| doctolib.de get-slots | 8 of 8 | 1 of 2 |
| doctolib.de search-practitioners | 9 of 9 | 2 of 2 |
| eventim.de | 9 of 9 | 0 of 9 |
| flixbus.de | 9 of 9 | 3 of 3 |
| idealo.de | 9 of 9 | 0 of 9 |
| immobilienscout24.de | 0 of 2 | 0 of 9 |
| jameda.de get-slots | 9 of 9 | 9 of 9 |
| jameda.de search-doctors | 9 of 9 | 9 of 9 |
| kleinanzeigen.de | 9 of 9 | 2 of 2 |
| lieferando.de | 9 of 9 | 0 of 9 |
| mobile.de | 5 of 9 | 0 of 9 |
| opentable.com | 9 of 9 | 0 of 9 |
| vinted.de | 9 of 9 | 7 of 8 |
| walmart.com | 9 of 9 | 0 of 9 |
| willhaben.at | 9 of 9 | 4 of 4 |
| zalando.de | 0 of 9 | 0 of 9 |
| zillow.com | 9 of 9 | 0 of 9 |
Where Wield failed
- zalando.de, 0 of 9. The run log shows a tab_error. The tab redirected from www.zalando.de to en.zalando.de, and the recipe's allowed origin needed an update. The HTML fallback has the same error in its 9 runs.
- immobilienscout24.de, 0 of 2. Both runs stopped with site_challenge, an AWS WAF bot check. Wield does not solve challenges. It stops and asks the user to solve the check in their own tab.
- mobile.de, 5 of 9. Three runs failed with a tab_error from a redirect, from suchen.mobile.de to www.mobile.de. One run failed with no error text.
- airbnb.com, 8 of 9. One run failed with no error text.
Some raw logs (eventim.de, vinted.de, willhaben.at and doctolib.de search-practitioners) hold a few more recipe runs than the summary, from timeouts and retries. Use the summary counts.
Where the agent failed
The agent had no passes on 10 tasks: bahn.de (1 run), eventim.de, idealo.de, immobilienscout24.de, lieferando.de, mobile.de, opentable.com, walmart.com, zalando.de and zillow.com. Most of these sites also blocked the agent in every run. The bot protection data covers that.
Two tasks had partial failures. On doctolib.de get-slots, 1 of 2 runs returned no items. On vinted.de, 7 of 8 runs passed, and the one failure was a navigation failure with no items.
Method and caveats
- Basis: 23 benchmark tasks on 21 sites. bahn.de is included, with 1 agent run. Zalando's HTML fallback is excluded because it reuses the main Zalando agent runs. immobilienscout24.de has 2 Wield runs.
- 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, in the user's everyday browser.
- Dates: 7 Nov 2026. The benchmark is still running, and the numbers will change.
FAQ
Which sites does Wield fail on?
zalando.de (0 of 9) and immobilienscout24.de (0 of 2) failed every run. mobile.de passed 5 of 9 runs, and airbnb.com passed 8 of 9. The causes are in the section above. The per-task pages for Zalando and ImmoScout24 have the raw runs.
Does a browser agent ever beat Wield in this benchmark?
Once. On airbnb.com the agent passed 4 of 4 runs, and Wield passed 8 of 9. On every other task, Wield passed at least as many runs as the agent.
Why do the two summary percentages differ?
The pooled counts (92% and 43%) use the 23 tasks and count every run. The summary (85.2% and 51.6%) uses 24 tasks, including Zalando's HTML fallback, and gives each task the same weight whatever its run count. Both come from the same runs.