Agents are only as capable as the environments they run in. That's the line Cursor pulled out again in its August 13 blog post—and this time paired it with a very practical solution: builds.
The Speed Math
Until now, every cloud session startup went through the classic three-step ritual: boot a machine, clone the repository, run installation scripts. For large, complex repos, the environment alone could take minutes before the agent could touch anything. Builds changes this: Cursor continuously prepares ready-to-use copies of your development environment in the background, running a build every hour by default, keeping warm snapshots where "repo cloned, dependencies installed, setup scripts already executed." When an agent launches, it forks a live machine instead of restoring from cold disk.
Internal numbers: environment startup is 10x faster, and time-to-first-token is 3x faster. Customers see the same results—Faire runs 2,000+ automated agent sessions per week, and its most complex monorepos now start in seconds. Better still, "a bad build can never take down the entire agent fleet."
As of August 17, builds is enabled by default for all new and existing environments, at no additional charge.
The Boundaries
Builds relies on filesystem snapshots, so two things need manual attention:
1. Installation commands should cover everything that can be pre-provisioned (e.g., dependencies) ahead of time. 2. Registries requiring private keys should use team or environment keys—user keys never enter snapshots and are only injected when an agent starts.
The start command still runs at your first prompt—it's reserved for services that must be launched fresh (e.g., spinning up Docker).
One more reliability win: cloud agents always start from the most recent successful build. If a dependency upgrade breaks your install script, that build simply never takes effect—you debug in the background while old and new sessions keep running normally.
It Rewrites "Waiting"
Cursor has turned "environment setup" from a per-session upfront tax into a continuously running background pipeline. For agent engineering, the real bottleneck was never how smart the model is—it's whether it can start working immediately. When startup drops from minutes to seconds, and a broken environment no longer punishes the whole fleet, handing more engineering work to autonomous agents stops being a slogan and becomes an executable default.