Superserve, a three-person Y Combinator startup, says the virtual machines it builds for AI agents are now running more than 100,000 machines a week. According to Superserve's company-submitted YC profile, teams including the AI assistant maker Lindy and Y Combinator itself use the service. Lindy's use is supported by public posts; independent confirmation that Y Combinator is a user was not found. The machines are pausable Firecracker microVMs, lightweight virtual machines that can be frozen and revived on demand.
The figure comes from Superserve's own YC company profile, published in October 2026. It is company-submitted, and no third-party audit turned up in the research for this article.
The problem Superserve is chasing is easy to state. Agents that work for days or weeks need a place to keep their files, running processes and memory, and containers or sandboxes tied to a single request were never designed for that. Superserve's bet is a dedicated virtual machine for each agent, one that sleeps at no cost and wakes when called.
What Superserve built
The company's tagline is "Virtual machines for long-horizon agents." According to its YC page, machines start from templates in under 50 milliseconds and can pause with memory and files intact, at zero compute cost. They can also run Docker and databases, the page says, along with full browsers.
Security is handled separately from the machine itself. Superserve says a broker holds API keys outside the VM, and outbound network traffic is blocked unless explicitly allowed. Its site also lists snapshot and fork features, an SDK and API, support for the Model Context Protocol (MCP), and compliance postures covering SOC2, HIPAA and GDPR. The product is open source and can be self-hosted, according to the company, which maintains a public GitHub organization called superserve-ai.
Who is behind it
Pavitra Bhalla is founder and chief executive. His YC profile lists time as a staff software engineer at Dosu, which builds code-reasoning agents used by 200,000 engineers, including teams at HashiCorp and Datadog as well as LlamaIndex. Earlier, he was a founding engineer and later director of engineering at Long-Term Stock Exchange (YC S17).
Superserve was founded in 2025. A founder post points to a pre-seed round from Open Core Ventures, though no amount has been disclosed. Dealroom, a startup data directory, lists a $125,000 YC investment in September 2026 and treats it as a standard batch deal. Dealroom is a compilation rather than an announcement, so the company has not confirmed that figure.
Named users and integrations
Lindy launched Teammate, an AI employee that lives inside Slack, on Aug. 10, 2026. Superserve's LinkedIn page tied itself to the launch: "If AI is going to work like an employee, it needs … a computer. Superserve powers the computers behind Lindy Teammates."
On June 12, 2026, Superserve announced an integration with Anthropic's Claude Managed Agents. The agent loop runs on Anthropic's side, while each tool call executes in a persistent Superserve sandbox. "The design decouples the brain from the hands," the company wrote in its announcement.
A post dated Oct. 1, 2026 described running Cursor Cloud Agents on Superserve, with durable state and self-hosted machines. A July 15 case study showed a Phaser Game Agent turning a one-sentence prompt into a playable game inside Superserve sandboxes.
A crowded field
Superserve is hardly alone in pitching microVMs for agents. Vercel made its Sandbox product generally available on Jan. 30, 2026, using per-request Firecracker microVMs and billing it as "the execution layer for agents." Support for Cursor Cloud Agents followed on Sept. 3, 2026.
Amazon Web Services announced Lambda MicroVMs on July 10, 2026, promising VM-level isolation and near-instant startup. Cloudflare Sandboxes and Akamai Agent Sandbox aim at the same job, and Akamai blocks outbound traffic by default. Baseten, meanwhile, has expanded into what it calls an "agentic cloud" after the team from Blaxel joined it.
Anthropic itself added self-hosted sandboxes and MCP tunnels to Managed Agents on May 19, 2026. Rogo, an institutional finance firm, was cited as an example of a customer running on Vercel Sandbox. OpenAI's Computer-Use tool takes a different route, running browsers in an environment OpenAI hosts.
Then there are the enterprise suites. Microsoft, SAP, Salesforce and ServiceNow each embed their own agent harnesses and governance tools, which gives them a natural way to bundle this kind of infrastructure for existing customers.
Pressure from security officials and regulators
The UK's National Cyber Security Centre said in August 2026 that organizations should always run AI agents inside a sandboxed environment, with controls on network access and resources.
Sandboxing is no guarantee, though. Tom's Hardware reported on Oct. 24, 2026 that red-team agents, working under lab conditions, built a covert channel through CPU cache timing that got around microVM sandboxes. The agents had no network and no shared files.
On the regulatory side, enforcement responsibilities under the EU AI Act took effect Aug. 2, 2026. Obligations for high-risk systems are staged through Dec. 2, 2027 and Aug. 2, 2028.
Signs of demand
Gartner's April 2026 Hype Cycle for Agentic AI found that 17% of organizations have deployed AI agents and more than 60% expect to. KPMG's third-quarter 2026 pulse survey put multi-agent systems in development or implementation at 25%, up from 6% earlier in the year. A second-half 2026 report from Nylas found 7.5% of respondents run fully autonomous "YOLO mode," compared with 4.2% in December.
On Aug. 10, 2026, the venture firm a16z wrote that computer-use infrastructure is shifting its focus from raw capability toward reliability, scale and return on investment.
Superserve did not disclose revenue, pricing or customer counts beyond the users it has named. Its next test is whether 100,000 weekly machines can hold up as larger rivals, with far deeper distribution, fold similar features into products customers already buy.
