Amplitude Accelerates Product Iterations with Wave

Amplitude introduces Wave, a product agent that analyzes data and replays, proposes technical plans, and manages implementation through code agents. The company claims shortened deployment cycles of 1 to 2 days and measurable gains, such as a 160% increase in messages via suggestion cards after customization.

Among beta clients, up to two-thirds of launches in one week

Companies from various sectors are using the beta version of Wave to identify and deploy improvements. Of all the experiments that lead to production, one in four is deployed on the same day the opportunity is detected, half in less than four days, and two-thirds within a week. Teams maintain control over automation: the majority send bug fixes directly to coding agents, while feature additions and major initiatives undergo human validation.

Compressed cycle and measured results on the homepage

Amplitude estimates that Wave enables improvements to be shipped in 1 to 2 days instead of about a month. On its homepage, the customization of suggestion cards is associated with a 160% increase in messages sent via these cards compared to the control after one day. The company highlights these results as indicators of a shortened iteration cycle and a rapid impact on engagement.

Wave architecture: data, coordination, and learning loops

Wave is the product agent developed by Amplitude. Teams define a scope, objectives, and metrics, while the agent gathers analytical data, session replays, user feedback, and code to detect and prioritize opportunities. It designs solutions, orchestrates their implementation with collaborators and coding agents, monitors performance, and integrates lessons learned into the next cycle. Product plans include a suggested solution, a technical plan, and acceptance criteria, which can be passed to coding agents, including Codex. Wave evaluates the impact of changes after deployment.

Case study: personalized suggestion cards after replay analysis

Jacob Newman describes Wave as a unified platform where agents sift through replays, analyses, and code to identify opportunities. In a sample of 100 replays, 80 users asked their own questions, and only 12 selected a suggestion, leading Wave to propose personalizing the cards with recent graphics, dashboards, or the original page. He illustrated the change and produced a technical plan that aimed to retain free input while switching to generic suggestions in the absence of context. After a brief review, the work was handed over to a coding agent and tested in A/B. Eric Carlson emphasizes that modest adjustments can encourage feature adoption.

Model distribution: reasoning, checks, and mockups

Amplitude states that Wave relies about 80% on OpenAI models, including GPT-6.1 Sol, GPT-6 Luna, and GPT-Image-2.5 Flare. GPT-6.1 Sol supports the opportunity manager and deep investigations into code and data. GPT-6 Luna performs frequent and low-cost tasks, such as code analysis and minor checks. GPT-Image-2.5 Flare generates product mockups from screenshots derived from replays.

Benchmarking: scores, latency, and inference costs

Amplitude claims to balance quality, latency, and cost to assign a model to each task, and Spenser Skates notes that OpenAI provides both powerful reasoning models and smaller, faster models. For the Opportunity Investigator, the evaluation focused on corroborating results, avoiding false positives, integrating feedback, recovering from tool failures, and aligning claims with evidence. GPT-6.1 Sol scored 4.06 out of 5 compared to 3.62, 3.52, and 2.94 for three alternatives, while displaying a median investigation time of 54 minutes. For the Data Science Agent, GPT-6.1 Sol scored 67.5 out of 100 compared to 69.3 for another leading model, with an average cost of $1.28 versus $2.63 per shipment, and better scores in compliance with tools, estimation and uncertainty, and anchoring evidence. Skates adds that the models can examine a problem, write and execute code, and then decide on the next step based on the results.

Stated ambition: towards self-improving products

Spenser Skates states that he aims, with OpenAI, for products that self-improve as a mode of software construction. Amplitude believes that as models advance in reasoning and coding, more investigation and implementation can be delegated to agents. The company is already observing agents that find and help deliver improvements in its own product while the team focuses on other priorities. Skates anticipates that teams will be surprised by the extent of what they will accomplish with software working alongside them.