Building cloud agents requires rethinking infrastructure, not just porting code
Original: What we’ve learned building cloud agents
Source: cursor.com ↗
Who: Posted on the Cursor blog, authored by Josh Ma. Cursor is the company behind the AI-assisted code editor of the same name; Josh Ma is an engineer on the Cursor team writing about lessons from their production cloud agents system.
What's new: Cursor's engineering team shares a year of hard-won lessons moving AI coding agents from running on a developer's laptop to running on their own dedicated servers in the cloud. The core finding is that this transition is not a simple lift-and-shift — it requires building what amounts to a full IT department for software agents, covering environment setup, long-running reliability, and state management.
How it works: The biggest quality factor turned out to be giving each cloud agent a complete development environment — compilers, credentials, network access, everything a human developer would have. Without this, agents fail silently, producing subtly worse code rather than crashing. For reliability, Cursor migrated from a fragile home-grown task system to , a dedicated tool for durable long-running work. They also separated three things that used to be tangled together: the agent's reasoning loop, the it runs on, and the conversation history — so any one of those can be replaced or restarted without losing the others.
The numbers: After migrating to Temporal, reliability crossed two nines — meaning the system is up and functional at least 99% of the time, up from a single nine (90%). Temporal now handles more than 50 million actions per day across more than 7 million unique workflows inside Cursor. Over 40% of Cursor's own internal are now created by cloud agents.
Why it matters: As AI agents take on longer, more complex tasks, the infrastructure wrapping them matters as much as the underlying model. Cursor's experience shows that environment setup and execution reliability are the ceiling on agent performance — and that the right answer is increasingly to give agents better tools and get out of their way, rather than hard-coding guardrails around them.