Engine B · AI Autopilot
Optimize live — without shipping a new build.
Expose the parameters you’re comfortable tuning, set their bounds, and let Autopilot run controlled experiments, personalized offers and difficulty adjustments. Every change is measured against a control group and rolled back if it hurts.
Capabilities
One policy engine, many levers
Remote config & flags
Typed parameters with developer min/max or allowed values, served via one SDK call.
A/B & multivariate tests
Deterministic, sticky assignment; significance tests; sample-ratio mismatch checks.
Bandits
Thompson sampling for offers and price points, with the control share held fixed.
Dynamic difficulty
Tune difficulty multipliers per experiment within bounds you set.
Personalized offers
Segment-targeted offers with per-user daily caps and holdout groups.
Notifications & re-engagement
Consent-gated messages with platform-wide frequency caps.
Rewards & missions
Churn-prevention rewards for high-risk users, measured against holdout.
Campaign optimization
Every campaign reports treated vs. holdout outcomes.
Control
You choose how autonomous it is
The default is approval required. Change it per project, any time.
01
Recommendation only
AI proposes. Humans implement. The AI cannot create live changes.
02Default
Approval required
AI drafts experiments and campaigns; a developer approves before anything starts.
03
Automatic within limits
Non-monetization parameters you authorized can start automatically inside your bounds.
04
Autonomous within authorized parameters
Winners can be promoted automatically — still bounded, still rolled back on harm.
Kill-switch
One toggle stops all AI-originated changes and campaign serving.
Automatic rollback
If a variant is significantly worse on the primary metric or a guardrail, it’s stopped.
Complete audit trail
Who or what changed which value, when, from what, to what — and why.
Safeguards
Optimization with ethics built in
- Control groups are mandatory and never below your policy minimum (default 10%)
- Only one experiment per parameter at a time to avoid contamination
- Experiments and personalization only for users who granted consent
- Child-directed projects: no personalized offers; monetization changes always need human approval
- Frequency caps on notifications and offers per user per day
Define a bounded parameter, read it in the app
// Dashboard / API: difficulty_multiplier
// type: number, min 0.8, max 1.2, autopilot_allowed: true
const m = aop.getParam("difficulty_multiplier", 1.0);
enemy.speed = baseSpeed * m; // exposure is recorded automaticallyHonest limits
- • Autopilot only changes parameters and content you explicitly expose — it never edits code.
- • Results need traffic: small apps may take weeks to reach statistical significance.
- • Reinforcement-learning policies are on the roadmap; today we use controlled experiments and bandits.
- • Webhook delivery is best-effort today; retries are on the roadmap.