Superserve

Superserve provides sandbox infrastructure for running AI agents in isolated virtual machines that can pause indefinitely and resume instantly. It's built on Firecracker microVMs, the same technology AWS uses for Lambda, but exposed as a developer-friendly API for teams building autonomous agents that need to run code, maintain state, and operate over long time horizons without arbitrary session limits. If you've tried to build an agent that needs to execute untrusted code or run for hours at a time and hit the walls of existing sandbox solutions, this is the infrastructure layer that's supposed to solve those problems. The pitch is durable, secure execution for agents that do more than one-shot tasks.

The core capability is durability. Most sandboxes give you a temporary environment that disappears when the session ends. Everything you built, every file you created, every state you accumulated goes away. Superserve lets you snapshot the entire VM state, fork it into parallel branches, and resume from any checkpoint later. This matters when you're building agents that need to explore multiple paths, backtrack from dead ends, or simply pick up where they left off after hours or days. There are no time limits on how long an agent can run, and the filesystem persists across sessions. For agents doing multi-step reasoning or long-running tasks like code generation and testing cycles, this persistence is essential. You can checkpoint before a risky operation and roll back if it fails.

Security is the other half of the pitch. Firecracker isolation means each sandbox is a proper VM, not just a container sharing a kernel with other workloads. The attack surface is much smaller, and the isolation guarantees are stronger. You get network egress controls so you can restrict what URLs an agent can reach, preventing accidental or malicious outbound connections. Credential brokering lets you provide API keys to the sandbox without exposing the actual values to the code running inside. Connection logging gives you an audit trail of what the sandbox tried to access. The platform is SOC 2 Type 2, HIPAA, and GDPR compliant, which makes it usable for teams in regulated industries who can't just spin up containers on a public cloud and hope nothing leaks. If your agents handle customer data or interact with production systems, these compliance certifications matter.

Under the hood, the developer experience is thoughtfully designed. You get real-time streaming of command output, so you can watch what your agent is doing without polling. File transfer in and out of the sandbox filesystem is built in, so you can seed the environment with data or extract results. Docker containers can run inside the VM if your agent needs that kind of flexibility. You can expose ports via public preview URLs, which is useful if your agent is standing up a web service that needs to be reachable from outside the sandbox. Custom templates let you pre-configure base images with your dependencies baked in, so agents don't waste time installing the same packages every run. Secrets management is handled separately from the code, so API keys don't leak into logs or snapshots. The developer experience is clearly shaped by people who have actually built agent systems.

The integration story targets the AI agent ecosystem directly. There's MCP compatibility, TypeScript and Python SDKs, and explicit support for Claude and other agent frameworks. The documentation includes guides for integrating sandboxes into agent harnesses and using them as execution environments for code-generating models. If you're building something that needs to execute arbitrary code in a safe environment, run for extended periods, and maintain state between runs, this is the infrastructure layer you'd otherwise have to build yourself. The alternative is cobbling together containers, persistence, and isolation from multiple services and hoping the seams don't create security holes. For teams that want to focus on agent logic rather than infrastructure, the abstraction is useful.

Pricing is pay-as-you-go with per-second billing. Compute runs about $0.05 per vCPU-hour, memory around $0.016 per GiB-hour, and storage is fractions of a cent. There's a free tier that doesn't require a credit card, so you can try it before committing. For teams that need predictable costs, the per-second billing keeps you from paying for idle time when agents pause. If you snapshot a sandbox and don't resume it for three days, you're paying storage but not compute during that window. The elastic resource allocation means you can scale up for intensive tasks and scale down when the agent is waiting, which matches how most agent workloads actually behave. The open-source nature of the underlying technology is worth noting, though the hosted service is the primary offering for most users who don't want to run their own Firecracker fleet.