1

Small, controlled changes

Tweak the agent in tiny ways: new intro line, slower pace, different scheduling flow.

2

Practice offline

Run against simulated leads and score replies with an automated judge—fast and cheap.

3

Shadow on real calls

Baseline speaks; the variant only proposes responses. Zero risk to revenue & compliance.

4

10% canary

If shadow results look good and compliance is clean, route ~10% of calls to the variant.

5

One objective + guardrails

Maximize (appointments × predicted show rate × compliance). Auto-rollback if it dips.

6

Promote winners

Winning variant becomes the new baseline. Then we iterate again.