Story
Cursor said the biggest lesson it has learned from building cloud agents is that the challenge is no longer just the model itself, but the entire environment surrounding it.
The company explained that cloud agents now operate inside dedicated virtual machines with their own development environments, network access, and the ability to handle long and complex tasks in parallel — independently from the user’s local computer.
According to Cursor, even a minor flaw in the environment does not necessarily crash the system or trigger obvious errors. Instead, it often appears as a subtle decline in output quality, making the environment itself part of the final product.
Details:
• Cursor said it effectively had to build what resembles a full IT department for agents, including network management, secret protection, and credential handling.
• The company explained that cloud agents need the ability to create pull requests, download dependencies, and conduct research autonomously.
• As agents increasingly operate for hours or days, reliability, interruption recovery, and system resilience have become core infrastructure challenges.
• Cursor migrated to the “Temporal” workflow platform after its earlier architecture proved too fragile for long-running autonomous systems.
• The company says Temporal now manages more than 50 million actions daily across over 7 million workflows.
• More than 40% of the company’s pull requests are now generated by cloud agents.
• Cursor separated “conversation state,” “machine state,” and the “agent loop,” allowing agents to move across environments and work asynchronously.
• The company says the next step is enabling agents to better understand their environments and repair problems autonomously instead of relying on constant human intervention.
What’s Next?
The race in artificial intelligence is no longer only about building stronger models, but about who can create the most stable and autonomous operating layer around them. As cloud agents continue to evolve, tech companies may gradually shift from building AI applications to building full operating systems for autonomous agents.
