Docker Launches Cloud Sandboxes for Long-Horizon AI Agents

AIDeveloper44 Team

Extending Agent Capabilities into the Cloud

Docker has announced the launch of Cloud Sandboxes, a service designed to support AI coding agents that perform long-horizon tasks. While Docker previously introduced local Sandboxes to provide microVM isolation for autonomous agents, the company identified a need for infrastructure that persists beyond the limitations of a local machine. As agents increasingly handle tasks lasting several hours—such as dependency migrations or large-scale refactoring—the requirement for stable, always-on compute has become critical.

The new Cloud Sandboxes offering utilizes the same microVM-based architecture as its local counterpart, ensuring consistent security protocols. Agents running in these environments are isolated within their own kernel and Docker daemon, protecting local files, networks, and secrets. By moving these workloads to Docker-managed cloud infrastructure, developers can continue their work after closing their laptops, or scale operations to run multiple agents in parallel without local hardware constraints.

Unified Workflow and Migration

A key feature of the Cloud Sandbox release is the ability to transfer tasks between local and cloud environments using a unified CLI. Developers can start an agent locally for quick iterations and move the process to the cloud with a single command, sbx move my-project --to cloud. According to Docker, this migration captures the sandbox’s filesystem and recreates it in the cloud environment, allowing for a seamless transition.

This unified approach is intended to simplify the development lifecycle. Instead of requiring different toolsets for local testing and cloud deployment, developers can use a consistent set of commands for both. This environment supports various popular coding agents, including Claude Code, Codex, Copilot, Antigravity, Open Code, and Hermes. Developers can also integrate their own custom tools using these pre-configured sandboxes.

Governance and Integration

Cloud Sandboxes include several features designed to assist with agent management and security:

• MCP Support: Users can connect Model Context Protocol (MCP) servers—such as Jira, Linear, or Grafana—once, enabling agents to access these tools whether they are running locally or in the cloud.
• Secrets Management: Sensitive information is stored once and injected via a proxy on a per-request basis, ensuring that agents do not have direct access to raw tokens or keys.
• Policy Definition: Developers can establish network policies to restrict which endpoints an agent can reach, a feature that will be further expanded with centralized governance for enterprise teams.

Pricing and Availability

Docker is providing a promotional $250 compute credit for users who sign up for the Docker Agentic Platform. This offer is valid for signups between September 22 and September 26, 2026. After the initial credit is applied, users are billed by the second for compute usage. Pricing tiers are organized by size, ranging from a Micro instance (1 vCPU, 2GB RAM) at $0.070 per hour to an XL instance (16 vCPU, 32GB RAM) at $1.118 per hour. The company notes that these costs cover sandbox compute, while model API usage and agent subscriptions remain separate billable items.

The introduction of Cloud Sandboxes reflects a broader shift in the development landscape toward long-horizon agentic workflows. By providing a persistent, isolated environment that operates independently of local power or connectivity, Docker aims to provide a more reliable foundation for automated coding tasks.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe