Over the past year, I've spent countless hours experimenting with AI agents—building prototypes, breaking things, and testing tools like Claude Code, Codex, and more. Each attempt taught me something new, but also surfaced the same frustrations: scripts that only ran on my laptop, fragile setups that didn't scale, no clear way to configure what each agent should do or which tools it could use, and a hard time getting agents to behave the way I wanted—let alone deliver amazing results. Too often I found myself battling environment quirks instead of exploring what agents could actually achieve. That's why I'm so excited about what we've been working on at Docker. We asked ourselves a simple question: what if running agents could be as easy, portable, and reliable as running containers? The result is cagent, a new runtime for AI agents, built to make experimentation simpler and collaboration easier - and today, we're open-sourcing it. Meet cagent cagent is an open source, Docker-native runtime designed to make agents first-class citizens in your developer workflow. Instead of fragile scripts or ad-hoc setups, cagent gives you a consistent way to define, run, and share agents using the same patterns you already know from Docker. At its core, cagent is a multi-agent runtime. You can define a single agent with a simple YAML file, or orchestrate an entire team of specialized agents that collaborate on tasks. Each agent can be configured with its own role, personality, and access to external tools. Supported providers Out of the box, cagent supports multiple model providers including OpenAI, Anthropic, Google Gemini, and others. You can switch between them easily through configuration, so you're not locked into a single vendor. Tools and MCP integration Agents can be given tools to extend their capabilities.

Read the full article →