Signal
The companies making software generation cheap are paying to own the layer that decides what survives, and now automating it. Datadog bought Eppo and shipped it as a product within a year, arguing that the faster teams ship, the more expensive it becomes to not know what is working. OpenAI paid over a billion dollars for Statsig, then kept the team and passed the software to Amplitude, whose chief executive put the thesis plainly: teams can generate more code than ever, and the bottleneck is knowing what to keep and what to roll back. At Kuaishou, agents now propose the experiments, write the code and judge the online tests themselves. When anyone can ship endless variants, the experiment pipeline, not the output, is the asset.
So what?
When making things is free, choosing what survives becomes the business. The experiment pipeline is the new factory floor.
Evidence
arXiv, 25 June 2026 (study): AgentX, deployed on the Kuaishou app, chains a Brainstorm Agent that turns past experiments, system architecture and data analysis into ranked proposals, a Developing Agent that implements them, and an Evaluation Agent that runs safe online rollouts with guardrail-vetoed A/B judgement. In three weeks three AgentX workers turned 374 ideas into 10 launchable rollouts, and the authors' framing is that innovation otherwise scales with headcount rather than compounding with evidence and compute.
Amplitude, 5 May 2026 (news): eight months after OpenAI bought the product testing startup Statsig for 1.1 billion dollars in stock, Amplitude took over the Statsig platform, brand and customers while the Statsig team stayed at OpenAI. Amplitude's chief executive: teams can generate more code than ever before, and the challenge is how to evaluate code before release, track what is working after it, and know what to roll back and when.
Datadog, 2 April 2026 (launch): Datadog Experiments became generally available, powered by the Eppo acquisition of May 2025. The release says AI has increased the pace and complexity of software releases exponentially, and its chief product officer Yanbing Li adds that the faster teams ship, the more expensive it becomes to not know what is working.
Worth asking
How much of your week goes into making things, and how much into finding out which of them actually work?
