The VMs Powering Mobile Agents (Instinct, Claude Code)
summary
AI agents are shifting from local execution to cloud-based virtual machines (VMs) to enable mobile access and better persistence. By leveraging Firecracker microVMs, platforms like Claude Code and Instinct create isolated, disposable environments that separate the agent's logic from its durable memory and identity. This architecture allows for rapid spin-up times and secure, scalable execution of tasks across different devices.
key points
- Claude Code uses a highly stripped-down Firecracker microVM with a custom Rust-based init process to minimize boot time and maximize isolation, keeping the agent's harness on the box while routing inference through a secure egress gateway.
- Instinct employs a 'sandbox-as-a-service' model via E2B, using a full Ubuntu environment where the agent's memory is stored as a Git repository of Markdown files in S3, allowing the agent to author its own persistent knowledge base.
- To handle authenticated web tasks, Instinct uses a server-side cloud browser pool with lease-based access, ensuring that sensitive cookies and credentials never actually touch the disposable execution sandbox.
who should read it
Software engineers and system architects interested in the infrastructure and security patterns used to build scalable AI agent platforms.
discussion positive
consensus
Commenters generally agree that Firecracker is the dominant runtime for agent sandboxes, with E2B being a primary provider. There is a collective interest in the lens of micro-VMs and the technical implementation of memory-in-git systems. The sentiment is largely curious and technical, though some users express security concerns regarding permissions and the potential for human-in-the-loop interventions.
top insight
The discovery that some advanced agent platforms may actually be using simple markdown files in git for memory rather than complex databases or graphs provides a grounding perspective on current AI agent architecture. This suggests that basic, version-controlled text files can be surprisingly effective for state management in early-stage agentic workflows.
Generated by an LLM from the article and thread. Check the sources before relying on it.