Define performance SLIs (e.g., p95 latency) and monitor them continuously. Add profiling/tracing for slow endpoints, use load tests for critical flows, and set budgets/alerts so regressions are caught early. Feature flags can help you roll back quickly if needed.