Virtual machines for long-horizon agents.

Introducing Claude Managed Agents on Superserve. Today, Superserve Inc. is announcing its collaboration with Anthropic to launch an integration between Claude Managed Agents and Superserve. The agent loop runs on Anthropic, and every tool call executes inside a persistent Superserve sandbox. That gives you full control over where the agent's work runs and what it can touch, without operating any agent infrastructure yourself. Superserve Inc. is building Superserve for a world where software is run by agents. That shift needs a new kind of cloud, one that's programmable to the core and built for agents that act on their own. Bringing Claude Managed Agents to Superserve was an obvious next step, because it's exactly the kind of agent Superserve Inc. built for. Claude Managed Agents is Anthropic's harness for long-running, asynchronous agents. Unlike interactive coding agents, it is built for work that runs unattended: multi-file refactors, support triage, data pipelines, overnight builds. The agent loop runs for as long as the task takes, including idle stretches that can last hours. The design decouples the brain from the hands. Anthropic runs the reasoning loop; every bash, read, and write executes inside a Firecracker microVM on Superserve, with tool results flowing back to the model so it can determine what to do next. Your execution environment, filesystem, and network egress stay under your control, with the snapshotting and egress controls production workloads actually need, and no infrastructure for you to manage. Start with its recipes: persistent dev environments, parallel sandboxed benchmarks, and pause-and-resume research agents, all ready to run. How it works. Anthropic runs the agent loop, model calls, prompt caching, compaction, session state, the event stream, web tools (web_search, web_fetch), and MCP server tools. Because MCP tools run on Anthropic's side, their credentials never reach your sandbox. Superserve runs filesystem and shell tools. bash, read, write, edit, glob, and grep execute inside a per-session sandbox on your account. You run one orchestrator that long-polls Anthropic's work queue, plus any apps that create sessions and stream events to your users. The orchestrator and your apps never talk to each other directly. Anthropic sits between them, routing work to the orchestrator and events to the apps. Scale either side independently. Your customer-facing app does not need to know where the sandbox lives or how it boots. Stateful sandboxes. Every Claude session gets a dedicated sandbox, and its filesystem persists for the life of that session. Between turns, sandboxes can be paused. Superserve snapshots the full VM state (memory, processes, and filesystem) at zero compute cost. When the next work item arrives, resume brings the same sandbox back in under 100ms with everything intact. You pay for compute only while the agent is actively working. This is where pause and resume earn their keep. Long-horizon agents spend most of a session waiting, not working: for the next turn, the next stage, the next ticket. Without it, you pay for a running VM the entire time the agent waits. The same snapshots let you branch. Fork a running sandbox and explore multiple paths from one shared state, each in its own microVM: run several approaches in parallel, then keep the one that works. Firecracker microVM isolation. Each sandbox is a Firecracker microVM, not a shared container. Process tree, filesystem, and network namespace are fully isolated. A compromised sandbox cannot reach another one, cannot reach your control plane, and cannot reach the internet beyond what you allow. The orchestrator creates every sandbox with a locked-down network: the runner can reach api.anthropic.com to post tool results, and nothing else. Need to clone from a private host or hit an internal API? Add the domain to the allowlist. Anything not listed is blocked. Your organization API key stays on the orchestrator host and never enters a sandbox. The orchestrator passes in only a scoped environment key, bound to the runner process, so the agent's own tool calls cannot read it. Phaser is building a complete game development agent with Claude Managed Agents and Superserve. The agents are given game design documents and then automate building the game code and assets in a secure sandbox environment, publishing the games ready for users to play. - Richard Davey, Founder and CTO of Phaser