Combine parallelism, selection and pruning. First make tests independent, then run them in parallel and shard across machines; the saving is close to linear only while tests do not compete for the same database or service. Add test-impact analysis or change-based selection so a small change does not run everything, with periodic full runs as a safety net. Then push checks down the pyramid, delete redundant high-level tests, replace UI login and setup with API or database setup, and quarantine flaky tests so retries do not inflate runtime. Balance shards by past run time rather than by file count.

Sources: Playwright sharding, Augment Code guide, Fowler, Playwright CI