The development
What happened
Amplitude describes Wave, a beta AI product agent that turns analytics, session replays, feedback and code into proposed improvements and implementation plans for coding agents such as Codex. Its 12 November OpenAI customer account reports a one-day homepage experiment and task-specific model trade-offs, rather than independently validated general productivity gains.
Key facts
Teams give Wave a product area, goals and metrics. It investigates opportunities, proposes changes with technical plans and acceptance criteria, coordinates work with people or coding agents, then tracks outcomes. Teams choose the level of automation.
Amplitude reports 160% more messages sent through personalized homepage suggestion cards than control after one day. This is card-driven sends, not total usage or revenue; the earlier 100-replay diagnostic sample is not identified as the full A/B-test sample.
In its separate Data Science Agent evaluation, GPT-6.1 Sol scored 67.5/100 versus 69.3 for another leading model, at $1.28 versus $2.63 mean model cost per dispatch. The lower cost accompanied a lower overall score; this is not an all-in product-cost comparison.
Amplitude estimates improvements can ship in 1–2 days. Among customer-launched experiments, flags and guides, it reports one quarter live the same day, half in under four days and two thirds within a week—a selected launched-work denominator, not every opportunity.
The context
Why it matters
The concrete workflow links AI-led investigation to an implementation plan and then a measured product change, with teams controlling what ships. The account’s different denominators matter: an experiment on card-driven messages, model-evaluation costs and launch times for work that actually shipped cannot establish a universal productivity or adoption improvement.