Experiment Manager — Concept
What the Experiment Manager is, at the product level.
Owner: @adminStatus: wipLast verified: 5/18/2026
Stub. The exemplar service doc will live here. To be co-written via
/qa-session experiment-manager.
Concept
The Experiment Manager is the heart of Alyson AI. It runs experiments that compare multiple agents against a client champion and grades them on chosen objectives to surface a winner with statistical confidence.
See the glossary for term definitions.
Why it exists
Today, picking which landing page / email template / send time / etc. to ship is a slow, manual process. The Experiment Manager turns this into a continuous, model-graded experiment loop.
Key abilities
- Run experiments with multiple agents, each with its own router strategy
- Grade agents using ML predictions (RPC, CVR) — not actual outcomes, because actual outcomes lag by weeks/months
- Auto-optimize agent splits over time, with the champion's weight locked
- Score using a composite of multiple weighted objectives
- Compute lift and statistical confidence against the champion
What it isn't
- A real-time analytics platform — it's a graded-experiment platform
- A replacement for end-of-period business reporting on actual revenue
Related
- Architecture overview
- Service docs: Experiment Manager service
- Agent grading deep-dive: TBD