Cursor Expands Enterprise Control with Dynamic Self-Hosted Cloud Agent Pools

Cursor has expanded its Self-Hosted Machines infrastructure for Cloud Agents, introducing dynamically scheduled worker pools, machine hibernation, and native support across multiple third-party sandbox runtimes. Under this updated architecture, Cursor maintains the agent planning, inference, and orchestration loops in its managed cloud, while delegating file edits, terminal operations, local Model Context Protocol (MCP) servers, and computer-use actions to customer-managed workers. These workers operate on an outbound-only HTTPS polling model, executing tasks on private VMs or sandbox providers—including AWS Lambda microVMs, Cloudflare, Daytona, Modal, and E2B—without requiring inbound network access into enterprise environments. This shift directly addresses the primary blocker for enterprise agent adoption: data gravity and compliance boundaries. Organizations running private source code repositories, strict VPC firewall rules, or bespoke build toolchains previously could not safely grant third-party SaaS environments arbitrary execution rights. Furthermore, workloads requiring specialized hardware—such as macOS runners for iOS pipelines or internal GPU clusters—were fundamentally incompatible with standardized SaaS runners. By enabling dynamic pooling across repositories and supporting idle machine hibernation with state snapshots, engineering organizations can provide secure, ephemeral compute for asynchronous AI agents without racking up continuous compute expenses. The architectural model reflects an established trajectory in DevOps infrastructure, directly mirroring the evolution of CI/CD runners like self-hosted GitHub Actions and GitLab Runners. As AI coding assistants transition from inline editor autocomplete into long-running autonomous systems that test, fix CI errors, and manage pull requests, the execution runtime must sit within the target infrastructure. Moving the compute layer on-premises or into dedicated virtual private clouds mirrors broader enterprise cloud-native patterns where control planes remain centralized while data and execution planes remain distributed and sovereign. In practice, DevOps and platform engineering teams should evaluate how to provision and lifecycle these worker pools within their existing infrastructure automation. Leveraging serverless microVM sandboxes provides instant spin-up and strict per-session isolation, preventing state pollution between agent runs. However, teams must recognize the operational trade-offs: self-hosted workers shift the burden of compute costs, image patching, and environment provisioning onto internal platform teams. Additionally, because tool outputs and execution transcripts still flow back to the orchestrator for downstream reasoning, governance teams must ensure that outbound data filtering aligns with organizational privacy and data residency policies.