Shipping a product improvement traditionally required roughly a month of manual scoping, engineering, and testing. Amplitude now estimates its proactive product agent Wave helps teams deploy those same updates in one to two days. This transition exposes a shift from labor-heavy product backlogs to algorithmic execution loops with automated task analysis. This framework challenges legacy roadmapping: an automated system identifies underperforming touchpoints directly from behavioral logs, synthesizes technical specs, and dispatches them straight to autonomous coders.
Yet enterprise leaders face a direct financial calculation: can autonomous agents coordinate code, visual mocks, and analytics without human teams repeatedly intervening to catch drift? The data suggests product planning is shifting from quarterly human debates to automated opportunity generation backed by measurable unit costs. Amplitude describes Wave as roughly 80% powered by OpenAI models, running an orchestrated tier of models that includes GPT-6.1 Sol, GPT-6 Luna, and GPT-Image-2.5 Flare to lower operational overhead while accelerating release velocity.
Automated Behavioral Audits and the 160% Lift
The mechanics of this agent-driven model surfaced during an internal audit of Amplitude's own homepage. The page presented generic suggestion cards designed to prompt conversations with an AI agent, but telemetry told a different story. In a sample of 100 session replays analyzed by Wave, 80 users typed their own questions while only 12 selected a suggestion card. Rather than leaving the feature to decay, Wave identified the behavioral disconnect and designed an automated fix.
Wave proposed personalizing the suggestion cards based on a user's recent charts, dashboards, or referring page, while generating technical criteria to preserve free typing and provide fallbacks. After human engineers gave the plan a brief review, they dispatched the tasks to coding agents, including Codex. In the experiment readout Amplitude shared, messages sent through homepage suggestion cards increased 160% relative to control after one day.
"The idea with Wave is you have a unified platform with agents that comb through massive amounts of session replays, analytics, and code to find opportunities to improve products," explains Jacob Newman, Principal AI Product Manager at Amplitude.
As Eric Carlson, Senior Principal AI Engineer at Amplitude, explained, the homepage test demonstrates how non-groundbreaking elements that are normally ignored can increase feature adoption in unexpected ways once systems actively look for friction.
Multi-Model Tiering and Unit Inference Economics
The sustainability of this autonomous pipeline hinges directly on compute unit economics. Deploying top-tier reasoning engines across every repetitive task drains engineering budgets; Amplitude addresses this by routing tasks through a tiered architecture. GPT-6.1 Sol supports the opportunity manager and deep investigations into code and data, while GPT-6 Luna handles frequent, lower-cost tasks such as scanning code and running small evaluation checks. For interface assets, GPT-Image-2.5 Flare generates product mockups using screenshots from session replays as visual references.
In Amplitude's Opportunity Investigator evaluation, GPT-6.1 Sol scored 4.06 out of 5 overall, compared with 3.62, 3.52, and 2.94 for three alternative models, while leading across all five investigation criteria. According to Amplitude's assessment, the model cost in this benchmark was $1.34.
"AI can now find useful work we never thought to assign it. OpenAI’s models can investigate a problem, write and run code, and use the results to decide what to try next."
Financial discipline appears in high-volume evaluations as well. In the Data Science Agent evaluation, GPT-6.1 Sol scored 67.5 out of 100 at a mean model cost of $1.28 per dispatch, delivering a 51% lower mean model cost per dispatch compared to another leading model that scored 69.3 at $2.63. In the interface evaluation layer, model tiering ensures that high-cost reasoning is strictly isolated to complex architectural tasks.
While marketing narratives promise end-to-end autonomy where software plans, codes, and ships itself without friction, the technical reality requires teams to keep human checks between agent proposals and live code dispatches. Amplitude may compress an estimated month-long development cycle into one to two days, yet the workflow still depends on brief human reviews before handing work to coding agents to ensure models adhere to evidence and avoid system failures.
How the picture shifted
Validating an interface hypothesis used to consume roughly a month of team reviews, whereas orchestrated models now ship the fix in two days. Slashing delivery cycles is compelling, yet the real leverage sits in the workload hierarchy: burning the $1.34 GPT-6.1 Sol runs across routine tasks drains budgets fast, so repository scans fall to the lighter Luna. The operational gain hinges entirely on confining high-cost reasoning models strictly to surfacing friction inside telemetry logs. Amplitude effectively replaces quarterly roadmap negotiations with an algorithmic loop tailored for interface optimization.