Answer engine summary
What is a multi-armed bandit for AI agents?
A multi-armed bandit is an experimentation policy that allocates traffic dynamically across variants based on accumulated evidence, instead of holding a fixed 50/50 split until significance. Zubby uses Thompson sampling — each variant maintains a Beta posterior over its conversion rate, the policy samples from each posterior at every assignment, and the highest sample wins that shopper. Winning variants accumulate more traffic in near real-time; losing variants keep a small exploration budget so the policy adapts when context shifts. We apply it across the AI surface: opener messages on the widget, recovery email subject lines, smart-offer tiers, and journey-branch choices. Independent bandits run per audience segment so paid-traffic winners don't bleed into organic experiments, and the dashboard surfaces conversion intervals plus revenue lift so you keep the statistical rigour without paying for it in lost weeks of equal exposure to losers.
Beta posteriors
Each arm tracks a Beta(α, β) distribution over its conversion rate.
Thompson policy
Sample from every posterior, pick the highest — exploration with provably good regret bounds.
Per-segment runs
Independent bandits per audience rule so cohorts don't cross-pollute.
Live confidence
Cumulative conversion, CI, winning-arm probability, revenue lift — all real-time.