Skip to content

Latest commit

 

History

History
29 lines (21 loc) · 1.66 KB

File metadata and controls

29 lines (21 loc) · 1.66 KB

Positioning

Where Agentic Sandbox fits relative to other agent runtimes — described by capability axis, not by vendor.

Design axes

Axis Agentic Sandbox
Hosting model Self-host only. No hosted control plane.
Isolation boundary KVM hypervisor (per-agent kernel). Rootless containers as a lighter alternative.
Session duration Designed for sessions measured in hours to days.
Data path All traffic stays on the operator's network.
Agent shape Bring-your-own agent. Claude Code is the primary tested agent; the runtime is agent-agnostic.
Orchestration Single-host today. Multi-host and Kubernetes are on the roadmap.

When this is a good fit

  • Agent workloads where source code or data cannot leave the network (regulated industries, on-prem, air-gapped).
  • Long-running sessions where the bottleneck is "did the terminal stay open" rather than cold-start latency.
  • Workloads where hypervisor-level isolation is required (untrusted code execution, adversarial workloads).
  • Internal platforms running agent workloads for multiple users on shared infrastructure.

When something else is a better fit

  • Short, transactional sandboxes for individual tool calls — a hosted microVM service will be cheaper and faster to start.
  • Container-based dev-environment-as-a-service workflows — use a dedicated dev-environment platform.
  • Turnkey hosted "agent as a product" with no self-hosting — pick a managed offering.

The shared theme: if the operational cost of running this yourself isn't justified by the data, isolation, or session-length requirements, a hosted service will be simpler.