Setup was the hidden tax

A long-running agent may spend its first minutes doing work that has little to do with the request: installing packages, warming caches, downloading browsers, and discovering the repository. Prepared builds shift that repeated setup into a reusable environment so the agent can begin with the code and tools already in place.

Freshness becomes a product decision

The speedup introduces a new question: when is an environment too old to trust? Lockfiles, system packages, generated clients, and repository bootstrap scripts can all invalidate a build. A dependable setup needs visible build history and a clear rule for rebuilding after important inputs change.

Measure the useful milestone

Time to first token can improve while total task time stays flat. The better measure is time to the first reviewable artifact: a plan, failing test, focused diff, or completed pull request. That keeps infrastructure optimizations tied to what developers actually experience.