The product manager's job just got surgically automated, from deciding what to build to measuring whether it worked.

The Summary

  • Amplitude launched Wave, an AI agent built on OpenAI models that autonomously identifies product improvements, ships code changes via coding agents, and tracks impact metrics — compressing cycles from weeks to 1-2 days
  • This isn't AI assistance; it's end-to-end product development automation where humans approve rather than execute
  • The playbook: let agents handle the full loop (insight → code → measurement) while product teams focus on strategy and approval gates

The Signal

Amplitude, the product analytics company, just turned its own infrastructure into a self-improving machine. Wave operates as a proactive product agent that doesn't wait for tickets or roadmap meetings. It scans usage data, identifies friction points or opportunities, then routes work to coding agents that implement fixes. The whole cycle, start to measurement, runs in 1-2 days instead of the industry standard of weeks.

This matters because it's the first real example of an agent handling the full product development stack. Not brainstorming. Not code review. The entire loop. Wave decides what's broken or underutilized, generates the fix, ships it, then tells you if it worked. The product manager becomes a governor, not a doer.

"Wave identifies what to build, helps teams ship changes with coding agents, and measures the impact."

The architecture runs primarily on OpenAI models, which means Amplitude is betting on frontier LLMs for both analytical reasoning (what to build) and code generation (how to build it). That's a different risk profile than companies using AI for support tickets or content generation. If the model hallucinates a feature priority or ships buggy code, users notice immediately. Amplitude is effectively dogfooding AI agents in the riskiest possible domain: their own product velocity.

Key shifts this enables:

  • Product orgs can test 10x more hypotheses per quarter without hiring
  • Engineering velocity decouples from headcount growth
  • The bottleneck moves from "do we have capacity" to "do we trust the agent's judgment"

What's not in the announcement: failure rates, guardrails, or how much human review still happens before deploy. Also missing: whether Wave works on greenfield features or just optimization of existing flows. The difference matters. Optimizing a checkout flow is narrow and measurable. Designing a new social feature requires taste, not just data. If Wave only handles the former, it's powerful but bounded. If it attempts the latter, we're in uncharted territory.

The Implication

Product teams should start mapping which decisions are truly strategic versus mechanically derivable from data. If an improvement can be spotted in analytics and fixed with code, an agent will do it faster than your sprint cycle. The humans who survive this shift will be the ones who can articulate why an agent's suggestion is wrong, or spot the opportunity it can't see in the qualitative fog. Start building that muscle now. And if you're running a B2B SaaS company, expect your competitors to adopt some version of this within 18 months. The cycle time advantage is too large to ignore.

Sources

OpenAI Blog