Mobile access is changing the architecture of AI agents. When an agent must keep working after a user leaves a laptop—and remain available from a phone—the provider needs to supply more than an interface to a model. It needs a controlled computer, a persistence layer and a way to safely reconnect the two.
An examination of Claude Code and Instinct environments by developer Rohan Adwankar illustrates two versions of that emerging stack. Both appear to use Firecracker microVMs, lightweight virtual machines designed for fast, isolated workloads. But they divide control, execution and memory differently.
Claude Code: a tightly managed tenant machine
Claude Code’s mobile-accessible environment appears to run in a dedicated Firecracker microVM with a custom Rust process as PID 1, rather than a conventional Linux init system. That process accepts host control over vsock, mounts disks and starts the coding harness.
The design separates customer state from platform software. A writable root disk persists across reclaim events, while the Claude harness, task launcher and skills are supplied on read-only volumes. When an idle VM is reclaimed, its processes disappear, but the persistent disk can be attached to a later boot. That explains how a session can feel continuous without keeping compute permanently allocated.
The security posture is equally notable. The environment has no inbound networking, outbound traffic is constrained through a gateway, and authentication is host-minted and rotated at boot. In practical terms, Claude Code is treating the runtime as a managed appliance: users get a durable workspace, while the vendor retains tight control over the operating surface and software supply chain.
For operators, this is an increasingly familiar cloud pattern applied to agents: isolate each tenant, make compute ephemeral, and preserve only the state needed to resume work.
Instinct: disposable execution, durable memory
Instinct appears to take a more modular route. Its agent sandbox identifies as an E2B environment, suggesting it uses a sandbox-as-a-service provider rather than operating the full VM fleet itself. The environment is a more conventional Ubuntu desktop with systemd, modest compute resources and a finite lifetime.
The strategic center of gravity is not the machine. It is `/memory`: a Git repository of Markdown files containing entities, workstreams, communication history and timelines. The repository is persisted to S3 as a Git bundle and updated by the agent. Its timeline reportedly condenses over time—from raw records toward hourly, daily and longer-term summaries.
That is a consequential product choice. Instead of making proprietary vector retrieval or a hidden memory API the primary interface, Instinct makes agent memory legible and versioned. A Git-and-Markdown memory store is inspectable, portable in principle and easier to reason about operationally than an opaque state system—though it also raises familiar questions about retention policy, repository growth, search quality and access controls.
The sandbox itself appears to function chiefly as an execution surface. Tool actions are routed to Instinct’s API, while short-lived cloud credentials support persistence. This separation lets the provider change models or backend orchestration without making the temporary VM the system of record.
Why the architecture matters
The move to mobile agents makes infrastructure decisions visible to buyers. Teams evaluating agents should now ask questions that once belonged mainly in cloud-platform reviews:
- **What persists when a session ends?** Workspace files, conversational context, task state and long-term memory may live in different systems.
- **Where does execution occur?** A dedicated VM, a shared sandbox and a server-side tool layer create distinct isolation and compliance profiles.
- **Who controls the runtime?** A sealed vendor-managed appliance can reduce tampering risk; a more standard environment may improve extensibility and observability.
- **How portable is memory?** File-based, versioned state can ease audits and migrations, but only if export and lifecycle controls are real.
- **What is the wake-up experience?** Fast cold starts and pause/resume behavior will directly shape whether phone-based interaction feels immediate.
What to watch next
The near-term competition will not be only over model quality. It will also be over reliable resume behavior, secure delegated access to real services, predictable cost per active user and governance for accumulated agent memory.
The important shift is simple: agents that follow users across devices are becoming cloud products with a conversational front end. The VM may be disposable, but identity, memory and control planes are becoming the durable moat.
