A Year of Cloud Agents: Why Environment and Durability Matter Most
How cloud agents evolved from ported local agents into a full operating layer, covering environment setup, Temporal-based durability, and harness design lessons.
What began a year ago as a simple port of local agents to the cloud has grown into a far more complex system: cloud agents now run on dedicated VMs, execute in parallel, and handle long-running unattended tasks. The team found that the single biggest driver of output quality is giving agents a complete, developer-grade environment — missing pieces rarely cause crashes but instead produce subtle quality degradation. This has pushed them to build infrastructure resembling enterprise IT for agents, covering VM provisioning, hibernation/resume, checkpointing, and network access controls.
On reliability, an early fragile work-stealing architecture was replaced by a migration to Temporal, which now handles over 50 million actions daily across more than 7 million workflows, with cloud agents generating over 40% of internal PRs. Decoupling the agent loop, machine state, and conversation state lets subagents run across different pods while clients correctly resync streamed output after retries.
The harness itself has evolved from tightly scripted, deterministic behavior toward giving smarter models more autonomous tools and decision-making power. Looking ahead, the goal is self-healing environments where agents can detect and resolve issues like missing secrets or blocked network access on their own.