Lessons Learned from Building Cloud Agents

The most critical factor determining the performance of cloud agents is providing a complete Development Environment equivalent to what a developer has. Unlike local agents, cloud environments require rebuilding VM infrastructure, dependencies, and network access from scratch, and if the environment is incomplete, the model's performance subtly degrades.

To ensure the reliability of long-running agents, Temporal was introduced to build a Durable Execution structure. This allows tasks to continue even during disruptions such as inference provider outages or node replacements, and currently cloud agents handle more than 40% of Cursor's PRs.

A Decoupling strategy is used to separately manage the agent loop, machine state, and conversation state. By running the agent loop on Temporal instead of the VM, the pod's lifecycle can be managed independently, and conversation state is efficiently handled through a separate storage and streaming layer.

As models advance, the role of the Harness that controls the agent is also changing. In the past, the harness worked by verifying and enforcing the agent's tasks, but it is now evolving toward handing over authority so that the agent can directly control tools and make judgments itself.