What we’ve learned building cloud agents

A year after the launch of cloud agents, their capabilities have significantly expanded from being simple extensions of local agents to operating on dedicated virtual machines with their own environments and dependencies. These advancements have introduced challenges such as environment setup, reliability, and orchestration, which are crucial for optimal cloud agent performance. Unlike local agents that benefit from the host’s development environment, cloud agents require a complete, independent setup, as any deficiencies can subtly affect their output. To address these challenges, a robust infrastructure including user tools, VM management, and network access controls has been built, resembling enterprise IT systems. Reliability has been enhanced by transitioning to Temporal for durable execution, allowing cloud agents to handle disruptions and long-running tasks effectively. The architecture now decouples agent loops from machine and conversation states, enabling flexible deployment across different pods and enhancing interaction with clients through an efficient append-only storage mechanism. As models have become more sophisticated, the harness controlling the agents has evolved to delegate more autonomy to the agents, reducing the need for manual oversight. Looking forward, the focus is on empowering agents with tools to recognize and self-correct environmental issues, thereby enhancing their autonomy and reliability further.