[KongchangAI]
· 1 min read· 938 words

Amplitude Builds Wave Agent with OpenAI: Shipping Product Improvements in 1–2 Days

Amplitude Builds Wave Agent with OpenAI: Shipping Product Improvements in 1–2 Days

Amplitude's Wave agent uses an identify-deliver-measure loop to cut product improvement cycles to 1–2 days.

Amplitude launched Wave, a product agent built on OpenAI models that aims to automate the full product improvement workflow. It proactively analyzes behavioral data to spot opportunities, uses coding agents to help implement changes, and tracks impact after launch to close the loop—compressing delivery to 1–2 days. The case highlights an "agent-calls-agent" nested toolchain and a path for vendors to tackle fuzzy decisions via data advantages, though the results are self-reported and boundary conditions remain undisclosed.

Product teams often get stuck in the same loop: collect feedback, prioritize, build, ship, analyze results—and weeks have passed by the time it's done. Amplitude's product agent Wave aims to break that cadence. According to its engineering team, Wave leverages OpenAI models to proactively identify what to build, help teams implement changes with coding agents, and measure impact—compressing the delivery cycle for product improvements to 1–2 days.

Wave product agent workflow diagram

What Wave Is: From Passive Analytics to Proactive Recommendations

Amplitude started as a product analytics company, and its traditional value has been telling teams "where users are dropping off." Wave goes a step further: it doesn't just present data—it proactively points out "what to build next."

Built on OpenAI models, the agent takes on part of the judgment work normally done by product managers—identifying opportunities in behavioral data and offering actionable improvement suggestions. Teams no longer need to hold a prioritization meeting first; Wave puts the candidates directly on the table.

Traditional product analytics tools are essentially a "data presentation layer"—visualizing event tracking, funnels, and retention curves for humans to interpret. The value of such tools depends on the analytical skill of the user, whereas a so-called "product agent" attempts to automate interpretation and decision recommendations as well: typically built on large language models, it connects to behavioral data and directly outputs conclusions about "what to do" rather than merely providing charts. The key to this shift is that the model needs to understand business context (metric definitions, user segments, historical changes), so vendors often rely on their accumulated domain data and prompt engineering to constrain output quality. Wave positions itself at exactly this layer, turning an analytics tool from a "query entry point" into a "recommendation entry point."

A Three-Step Loop: Identify, Deliver, Measure

According to Amplitude's description, Wave's workflow revolves around three stages.

Identify: Analyze product usage data to find improvement directions worth investing in. This step replaces manually scanning dashboards and piecing together insights.

Deliver: Wave helps teams implement changes with coding agents. The gap between product recommendations and code implementation is bridged by the AI toolchain, reducing the communication overhead from "wanting to do it" to "getting it done."

Measure: After changes ship, Wave continues to track impact, forming a data loop. This means the next recommendation can be adjusted based on real results from the previous round, rather than iterating on gut feel.

It's the three steps chained together that produce the 1–2 day delivery cycle. No single step is novel on its own, but putting recommendation, implementation, and validation into the same pipeline is what truly compresses waiting and coordination time.

A "coding agent" refers to an AI programming tool capable of autonomously reading a codebase, editing files, running tests, and iterating on errors—representative forms include LLM-based IDE plugins and command-line agents. Unlike code completion, a coding agent receives task-level descriptions (e.g., "add event tracking to this button") and locates the relevant files itself to make changes. In Wave's workflow, product recommendations must be converted into such task descriptions before a coding agent can execute them, so the granularity and executability of the recommendation directly determine the success of the delivery stage—this is also the most fragile link in an "agent-calls-agent" structure: when the upper-layer agent gives vague recommendations, the lower-layer coding agent may produce changes that deviate from expectations.

Why This Case Is Worth Watching

Product agents are a relatively difficult direction in AI applications. Writing code and handling customer service have clear inputs and outputs, whereas "deciding which product features to build" involves a great deal of fuzzy judgment and context. Amplitude chose to enter through its data advantage: it holds behavioral data from customers' products, on which the model can make specific recommendations rather than generic ones.

Another signal is the nesting of the AI toolchain—Wave calls coding agents to complete tasks while itself serving as the team's product assistant. This "agent-calls-agent" structure is becoming a common form for deploying AI applications.

It should be noted that, based on public information, these are the vendor's self-reported results. The original text does not elaborate on how broad a range of requirements the 1–2 day cycle covers or what team sizes it applies to. For product teams evaluating whether to adopt a similar approach, these boundary conditions are more worth probing than the speed figure itself.

Implications for Product Teams

If your team still relies on the old three-part formula of "data dashboard + requirements doc + scheduling meeting," Wave's approach offers a reference point: let AI do the filtering and initial judgment, and let humans make the final trade-offs. What's truly scarce isn't analytical capability but the push that moves analytical conclusions all the way to launch—and that's exactly where agents are well suited to fill in.

Background Notes

The "1–2 day delivery cycle" repeated throughout the article needs to be understood in the context of the normal cadence of product improvement. The traditional process includes requirements gathering, prioritization meetings, design reviews, development scheduling, testing and launch, and impact analysis—much of the time is spent on cross-role waiting and information alignment rather than actual coding. The essence of compressing the cycle is reducing serial waiting: making recommendation, implementation, and validation parallel or automatically connected wherever possible. Note, however, that this figure typically applies only to small-grained changes (e.g., copy adjustments, experiment parameter changes, localized interaction improvements); requirements involving architectural changes or multi-team dependencies are still hard to accelerate this way. When evaluating such solutions, focus on the distribution of requirement types they cover rather than the time taken by a single best-case scenario.

#Amplitude Wave#OpenAI#product agent#coding agent#product analytics#AI agent workflow
Share:

Related articles