How to keep AI agent traffic polite to websites

#The limits

Every run stays inside these limits. The interpreter and the extension enforce them in code, so a recipe cannot exceed them by mistake.

Reviewers check the same budget before a recipe is published. A recipe with more than 15 requests per run gets a warning, and one with more than 40 fails the check. Loops of more than five requests must set a delay between calls.

#Why pacing protects you

Bot protection watches for bursts. A site that sees many fast requests from one browser can answer with a check or block the session. A recipe that paces its calls looks like one person using the site, which is what the site's own pages do.

When a site does show a check, Wield stops with the error site_challenge and asks you to solve it in your own tab. Wield does not solve or skip these checks.

Pacing also protects the site. Recipes are built for one lookup at a time that you start yourself, not for polling in a loop. Per-site pacing is shared by every client on your computer, so several agents against one site share one pace.

#How the benchmark was paced

The benchmark ran three fixed queries per task, three times per side. The plan runs in rounds. Run 1 of every query goes first, then run 2, then run 3. If a budget stops the run early, every query still has at least one result.

The pace was about 30 site requests per hour. After a site blocked the browser agent, the pacing rules kept the agent away from that site for the rest of the run, so that site has fewer runs than planned. The full method is on the benchmark page.

#What you can do

• Ask for one lookup at a time, with the dates you need.
• Use the smallest date range that answers the question.
• Do not loop a recipe to watch for changes, such as cancellations.
• Read the recipe's terms note on the site page, and follow the site's own rules.
• If you see a site check, solve it yourself in your own tab, then ask again.

#FAQ

#Why does Wield limit requests if the site allows more?

The limits keep each run small and in line with one person's session. A recipe that needs more than 40 requests fails review, so a run cannot turn into a crawl. Wield is built for lookups that you start yourself.

#Will a site still block Wield?

Sometimes. In our benchmark, Wield recipes passed most tasks, and some sites failed for Wield as well. The Zalando recipe passed 0 of 9 runs, and the mobile.de recipe passed 5 of 9. When a site shows a check, Wield stops and tells you.

#What does limit_exceeded mean?

A run hit one of its limits: steps, requests, response size or time. Narrow the query, or ask for a smaller date range, and run it again.

#Do several agents share one site budget?

Yes. Pacing is set per site across all clients on your computer, so parallel runs against the same site share one pace.