Superserve launches at YC with microVMs for long-running coding agents

A coding agent fixes a bug all afternoon, then the user leaves. The next day the agent has to rebuild the whole working set to pick up again. Superserve launched on Y Combinator today to remove that step. The company says it already runs more than 100,000 machines a week for other teams. Named customers include Lindy and Y Combinator itself.

Each session gets its own virtual machine with the code, dependencies, database, and browser already set up. The machine can be paused with no compute charges, then resumed with files, memory, and processes intact. Agents can also fork a paused machine. That lets one try two different fixes from the same starting point.

What the launch states

Superserve is in the Fall 2026 YC batch. The launch page records 5 votes at the time of writing. Founder Pavitra said in the post that each machine is a microVM with its own kernel. The user controls the network rules, which services the agent sees, and which third-party keys reach it. Credentials are brokered outside the virtual machine, so the agent itself never holds the raw key.

The post names three operations an agent cares about: pause, resume, and fork. Pause keeps state while work waits. Resume continues from where the agent stopped. Fork makes a separate copy of a paused machine, with its files, memory, and processes. The company says agents can use fork to run evals or reinforcement-learning rollouts from one base state. Each copy keeps its work separate from the others.

The speed numbers, as Superserve states them

Superserve cites a benchmark by ComputeSDK that it says it won. The test measured time from create to the first successful command, with 100 machines started at once. Superserve says its median was 0.37 seconds. The company's launch page claims it beat E2B, Daytona, Modal, Vercel, Cloudflare, and Google Cloud Run on that test.

Its own benchmark page gives slightly different numbers. On the ComputeSDK test, Superserve lists its median as 0.41 seconds and Cloudflare's as 0.68 seconds. Google Cloud Run on the same test is 3.12 seconds, Superserve says. The company also cites a separate test called AI Sandbox Bench. On that test it lists a cold-start median of 0.52 seconds and a snapshot-start median of 0.55 seconds.

One customer, which Superserve does not name, measured a cost drop against its earlier setup. The customer said Superserve was 77 percent cheaper than its custom-built sandbox on Amazon EC2. The launch post gives this as the customer's own figure.

Named customers and the team

Lindy founder Flo Crivello is quoted on the Superserve site. He says Lindy benchmarked several sandbox providers and picked Superserve for speed and reliability. Phaser creator Richard Davey is also quoted. He says Superserve lets his team give every user a game studio. The studio comes with the Phaser engine, Playwright, and build tools. His team does not have to run the infrastructure.

Founder Pavitra was a founding staff engineer at Dosu. He built code-reasoning agents used by 200,000 engineers there, at firms including HashiCorp, Datadog, and LlamaIndex. Co-founder Alejandro led the Kubernetes platform at Oracle. Co-founder Amit built the copy-on-write storage engine in Superserve's Firecracker fork. The company says that engine lets machines share one template while keeping each agent's work apart.

Source

Y Combinator launch page for Superserve. Benchmarks on Superserve's own site.