Cursor, the AI coding tool built on top of VS Code, said it will start preparing cloud agent development environments continuously in the background instead of assembling them from scratch every time an agent starts. The company calls the resulting snapshots “builds,” and plans to turn the feature on by default for every environment on August 17.
Previously, a cloud agent session first had to spin up a machine, pull down the codebase, and execute setup scripts before an agent could do anything useful, a process that could run several minutes on a large repo. Builds instead refresh every hour by cloning the repository and installing dependencies ahead of time, so a new agent forks a warm, already-configured machine rather than assembling one from a cold start.
Cursor says the change gets a response from an agent up to three times faster, and that its own internal environments now boot ten times faster with a roughly threefold gain in time to first token. Both figures are Cursor’s own measurements of its own product, not an independent benchmark, and they describe how quickly an agent begins responding to a prompt, not how quickly it finishes the coding task itself.
The more consequential design choice is what happens when a build breaks. If a dependency bump or a broken container config causes the environment setup to fail, that build is discarded and never goes live, and agents keep running on the last version that worked. A developer can debug the failure, manually or with another agent’s help, while existing and new sessions continue uninterrupted on the older, known-good snapshot.
Faire, an e-commerce marketplace and Cursor customer, said its agents now complete more than 2,000 automated runs a week with no manual prompting; senior engineer Blair McAlpine said broken builds no longer take down the company’s fleet.
That fallback has a cost Cursor does not dwell on. An agent can spend an entire session working against a build that is, by definition, out of date relative to whatever branch a human is actively fixing. Decoupling the agent from a broken working state is genuinely useful once agent runs are frequent enough that one bad commit would otherwise stall a whole fleet, but it trades that stability for staleness risk. Cursor’s partial answer is a configurable threshold that stops an agent from starting too far behind the default branch, plus a dashboard that ties every agent run to the exact build and commit SHA it used, so a stale run is at least traceable after the fact.
Environment preparation is not a feature developers request by name. It is the plumbing that decides whether an agent starts working in seconds or minutes, and the fact that a coding-agent vendor is investing engineering effort here is itself informative: for this category, the competitive edge is shifting away from raw model quality and toward the infrastructure surrounding the agent, where Cursor can afford to run a fresh build every hour at no extra charge.
Teams running large or complex repositories through Cursor’s cloud agents should check that their install commands front-load anything cacheable, since the August 17 default rollout applies the same hourly build cycle regardless of repo size, and a poorly structured install script will blunt most of the speed gain.
Cursor described the builds feature in a post on its own blog, published in August 2026.
