Agent feedback loop
Every 4 hours the loop collects the bot's failed drafts, customers who had to write again, errors and reviewer decisions. Claude Opus 5.5 finds the causes and fixes them in a sandbox. New rules go live only if the safety tests pass and a replay of the day's failures scores at least as well. Code changes wait here for approval.
Last run
Loading…
Trend (per day)
Causes found
Loading…
Shipped changes, watched on live traffic
After a change goes live, every run compares the replies handled since with the same span before. If the judge pass rate, errors or frustrated customers got worse, the change is rolled back on its own.
Loading…
Code changes waiting for approval
Loading…
Live lessons
Rules added to each brand's brain by the loop. The drafter and the judge both read them.
Loading…
Store training (past tickets)
Knowledge mined from a store's own support history, and how the bot scores against what that team really answered on held-out cases.
Loading…
Today's evidence
Loading…