agenthog-experimentlisted

# Run an experiment with AgentHog An experiment is an analysis window on a multivariate feature flag. The flag delivers variants (deterministic per-user hashing in the SDK — no assignment service); the experiment compares treatments against a control on a metric and reports each arm's chance to beat it. Everything runs through the `ah` CLI; every mutation needs a write-scope token. Follow the phases in order. Each ends with a check — do not proceed past a failed check. ## 0. Preconditions - The AgentHog SDK is integrated and verified in the target app (events arrive in `ah events`). If not: stop and run the **agenthog-integrate** skill first. - `ah` is authenticated with **write** scope: `ah whoami` shows the project and `scope: write`. If read-only, ask the user to run `ah login --write`. - You know the **decision metric**: one event name that already exists (check `ah events top`) or a goal (`ah goals list`). If the metric event doesn't exist yet, instrument it and verify it arrives before doing anything else. ## 1. Create the flag ``` ah flags create checkout_cta --variants control:50,b:50 --desc "new checkout CTA copy" ``` - Key: lowercase `[a-z0-9_-]`, ≤64 chars. Variant names the same. One arm should be named `control` unless the user says otherwise. - A/B/C: `--variants control:34,b:33,c:33`. Percentage rollout of a non-experiment feature: `ah flags create new_nav` (boolean) then `ah flags rollout new_nav 10`. - Start experiments at full traffic (`--