Cursor Builds: Reimagining Cloud Agent Startup with Pre-warmed Environment Snapshots

Background

In its August 13 blog post, Cursor returned to a familiar line—*"Agents are only as capable as the environments they run in"*—and offered a concrete fix called builds.

The speed problem

Previously, every cloud agent session had to walk through three steps: boot a machine, clone the repository, run install scripts. On large, complex repositories, the environment alone could take several minutes before the agent could do anything useful.

Builds change the model:

• Cursor continuously prepares ready-to-use copies of the development environment in the background.
• By default, a fresh build runs roughly every hour.
• Each snapshot has the repository cloned, dependencies installed, and install scripts already executed.
• When an agent starts, it forks from a warm machine rather than restoring from cold disk.

Internal numbers

• Environment startup: ~10x faster
• Time-to-first-token: ~3x faster

Customer results

• Faire runs 2,000+ automated agent sessions every week.
• The most complex monorepos now start in seconds.
• A broken build can never take down an entire agent fleet, because sessions only fork from the last successful build.

Availability

As of August 17, builds are enabled by default for all new and existing environments, at no additional cost.

Boundaries and configuration

Because builds rely on filesystem snapshots, two things need manual handling:

1. Install commands should pre-install everything that can be pre-installed (for example, dependencies). Anything left in the install step runs on every fresh build, which slows things down. 2. Private registries requiring keys should use team- or environment-level keys. User-level keys are not baked into the snapshot; they are injected only when the agent starts.

The start command still runs on the first prompt of a session. It is reserved for services that genuinely need a fresh boot (for example, launching a Docker daemon).

Resilience model

Cloud agents always launch from the most recent successful build:

• If a dependency upgrade breaks the install script, that build simply does not promote.
• Older sessions keep running on the last good build.
• Engineers can debug in the background without disrupting active agents.

Why it matters

Cursor reframes "environment preparation" from a per-session tax into a continuously running pipeline. For agent engineering, the real bottleneck has never been how smart the model is, but whether the agent can immediately get to work. When startup drops from minutes to seconds, and bad environments no longer cascade across a fleet, handing more engineering work over to autonomous agents becomes a practical default rather than a slogan.

Key points

• Hourly pre-warmed snapshots replace cold-boot + clone + install on every session.
• ~10x faster environment startup, ~3x faster time-to-first-token in Cursor's internal benchmarks.
• Always forks from the last successful build, isolating failure to individual broken builds.
• No extra cost; default-on for all environments since August 17.
• Install commands must cover everything pre-installable; private registry keys should be team/env-level, not user-level.
• start command still runs per session for services requiring a truly fresh boot.

Reference

• Cursor blog post: August 13, 2025 (feature announcement for builds).
• General availability date: August 17, 2025.