Deploying a multi-agent system requires strict runtime isolation. XAgent addresses this need by utilizing a dynamic planning engine and orchestrating complex LLM workloads within a secure Docker sandbox. This approach safeguards your host system from untrusted, dynamically generated code or commands executed by the autonomous agents. Core Architecture Elements
XAgent Engine: Executes the core task loops. It handles runtime task decomposition, tool selection, and communication.
Docker Sandboxes: Provides a secure microVM boundary with an independent kernel and filesystem. This isolates agent executions entirely from your host operating system.
Dynamic Orchestration: Connects to proprietary or local models (via integrations like Xinference) to process instructions dynamically. Step-by-Step Deployment Guide 1. Setup Your Local Workspace
Clone the official platform repository and copy the environment template to configure your external keys.
git clone https://github.com/xorbitsai/xagent.git cd xagent cp example.env .env Use code with caution. 2. Configure Model Credentials
Open the .env file. Populate it with your required provider credentials (such as your OpenAI, Anthropic, or local model endpoints). This allows XAgent to safely interact with model APIs. 3. Initialize the Multi-Agent Framework
Launch the full platform in detached mode using Docker Compose. This containerizes the web control plane and local execution stack. docker compose up -d Use code with caution. 4. Configure Admin Access
Open http://localhost:80 in your web browser. You will be automatically redirected to /setup to create your master administrator account. If you ever need to reset these credentials later, run the CLI utility directly inside the app scope:
python -m xagent.web.reset_admin_password –username Use code with caution. 5. Provision the Secure Agent Sandbox
To ensure agent tools operate under an isolation layer, utilize Docker Sandboxes to configure a distinct container execution workflow:
# Provision a dedicated agent sandbox bound to your current workspace folder docker sandbox create agent ./path/to/workspace docker sandbox run agent ./path/to/workspace Use code with caution.
Alternatively, for multi-agent workflows requiring custom tools or shared volumes, launch a detached sandbox and attach your execution script via configuration manifests or manual commands:
docker sandbox run -d –name my-agent-sandbox claude docker cp ./custom_agent_binaries my-agent-sandbox:/usr/bin/ Use code with caution. Security Best Practices Building AI Teams with Docker Sandboxes & Docker Agent