Amplitude via Cursor: risk-tiered agent review in a customer-feedback pipeline

Source: Cursor customer story quoting Amplitude CTO Curtis Liu, staff engineer Adam Lohner and engineering head Spencer Pauly, April 15, 2026. A vendor-written customer case, with named participant testimony but no detailed audit or matched control.

One organization, several interconnected stages. Field teams send customer bug reports and requests in Slack. An automation checks Linear for duplicates, appends new customer context if ticket exists, otherwise explores codebase, creates ticket and proposes PR. Engineers use cloud agents for prolonged work and bring them local for closer iteration; separate hourly automations migrate CSS and legacy React. Cursor Bugbot reviews proposed code; another agent classifies PR risk. Amplitude says around 60–70% of low-risk PRs merge without additional developer work, whereas higher-risk changes route to engineers. Critical denominator: this is not 60–70% of all PRs, nor a zero-human-accountability declaration. The source does not specify the risk classifier’s rubric, confidence checks or whether that subset includes customer-originated changes. Deployment without developer involvement is stated as a next goal, not current achieved practice.

Output vs outcome: 3× weekly production commits after cloud agents and 1,000+ unprompted agent runs/week are self-reported volume metrics. No initial vs final customer-request examples, regression/incident rate, human review minutes, percent of low-risk across all PRs, per-change spend, launch impact or 30-day maintenance. The mention of “low-risk” requires operational boundary and downstream verification before treating waiver of human diff review as safe. Compare Linear’s human approval and CX notification, review-autonomy map, and single-feature trace question.