Superserve (YC F26) gives each AI agent session its own cloud virtual machine, with the application environment it needs to build, run, and test software. It is for teams running coding agents, autonomous workers, and evaluation environments whose work now lasts hours, days, or weeks rather than a single turn. A session can start from a template that already holds the repo, dependencies, services, and a browser, then keep files, memory, and processes between turns, the layer many of those teams have been assembling themselves on ordinary cloud VMs. A machine can pause and resume later, and paused time is not billed as compute. Teams can snapshot a running VM and fork it so several attempts run in parallel from the same state. Each sandbox is a Firecracker microVM with its own kernel. A credentials broker adds API keys only on outbound requests to hosts the team allows, so the secret stays outside the sandbox. The company site cites a ComputeSDK burst test from October 23, 2026, with a 0.41 second median from create to the first command when 100 sandboxes start at once, ahead of Cloudflare, Vercel, Modal, Daytona, E2B, and Google Cloud Run in that run. The company says it already runs more than 100,000 machines a week for teams including Lindy and Y Combinator. YC visiting partner Max Kolysh wrote that YC's own agent harness runs on Superserve, along with Lindy and more than 50 other teams. Founder and CEO Pavitra Bhalla previously built code-reasoning agents at Dosu, used by 200,000 engineers including teams at HashiCorp, Datadog, and LlamaIndex, and was a founding engineer and director of engineering at Long-Term Stock Exchange (YC S17). He is building Superserve with Alejandro, who led the Kubernetes platform at Oracle, and Amit, who built the copy-on-write storage engine in their Firecracker fork.