What we’ve learned building cloud agents

2026-05-21 · Source: Cursor Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Software Development & Engineering · Depth: Intermediate, medium

Summary

Cloud agents, launched a year ago, have evolved from simple local agent extensions to dedicated virtual machines with isolated environments, dependencies, and network access, enabling parallel, unattended, and longer tasks. This shift introduced challenges in environment setup, reliability, and orchestration. Key lessons include the critical importance of a full development environment for agent output quality, necessitating infrastructure for user tools, VM hibernation/resumption, and checkpointing. Durable execution is vital for long-running tasks; migrating to Temporal improved reliability from one 9 to over two 9s, now handling 50 million actions daily across 7 million workflows. The architecture also benefits from decoupling the agent loop, machine state, and conversation state for independent management and efficient streaming. The agent harness design is moving towards empowering agents with tools, allowing them to control tasks like multi-repo setups and CI Autofix. Future focus includes self-healing environments, such as "autoinstall," for agents to report and resolve issues like missing secrets or blocked network access.

Key takeaway

For AI Engineers or MLOps teams building or deploying cloud agents, recognize that robust environment provisioning and durable execution are paramount. Your agent's performance directly correlates with its access to a complete development environment, not just the model's intelligence. Consider adopting durable execution frameworks like Temporal to manage long-running tasks and ensure reliability, moving beyond simple local agent porting. Decouple agent, machine, and conversation states to enhance scalability and maintainability. Empower agents with tools to self-manage tasks, reducing reliance on hardcoded harness logic.

Key insights

Cloud agent effectiveness hinges on robust development environments and durable, decoupled execution, shifting from local extensions to an operating layer.

Principles

• A full development environment is crucial for cloud agent output quality.
• Durable execution primitives are essential for long-running cloud agent tasks.
• Decouple agent loop, machine state, and conversation state for flexibility.

Method

Migrate from work-stealing architectures to durable execution platforms like Temporal, structuring workflows into shorter, task-specific units and splitting activities for better timeout/retry handling.

In practice

• Reconstruct full dev environments for cloud agents, including enterprise IT infrastructure.
• Implement Temporal for agent workflow durability and reliability.
• Provide agents with tools (e.g., GitHub CLI) instead of hardcoded harness logic.

Topics

• Cloud Agents
• Development Environments
• Durable Execution
• Temporal Workflows
• Agent Orchestration
• Self-healing Agents

Best for: AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• The Age of Async Agents — Cognition's Walden Yan & OpenInspect's Cole Murray — Autonomous background agents, powered by advanced models, demand robust infrastructure and careful architectural choices for security and scalability.
• Long-Running Agents — Long-running agents achieve sustained progress by externalizing state and decoupling components, overcoming single-session AI limitations.
• Building AI Agents That Survive Production — Production-ready AI agents must be designed for dynamic execution, durability, and defensibility to tolerate inevitable failures.
• Impressions from visiting OpenAI, Anthropic, & Cursor — Cloud-based AI agents are becoming the standard for long-running tasks, driven by improved models and infrastructure.
• Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems [R] — Agent performance degrades longitudinally, with memory policy impacting lifespan more than base model upgrades.
• Long-running Agents — Long-running agents overcome AI limitations by externalizing state, enabling sustained, multi-session progress on complex tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Cursor Blog.