Over the past two years, the AI Agent wave has swept through the entire developer community.
From the early days when AutoGen put the concept of “multi-agent collaboration” on the table, to the subsequent launch of various Agent orchestration solutions by major frameworks, and then to the 2025 MCP protocol that enabled Agents to truly connect to external tools, the entire industry has been racing in the same direction: making AI not just chat, but truly “get work done.”
It sounds flexible, but in actual use there is always a sense of “running naked”—the Agent operates on your files, accesses your browser, and calls your APIs, with every step being a black box. You can only find out what it did by checking the logs afterward, and often the logs are not detailed enough.
This is the core pain point that the CopilotKit team set out to solve when launching OpenBot.
Their approach is quite interesting: since we cannot trust an Agent’s “self-discipline,” why not give every Agent its own dedicated “computer”—a truly independent container environment with its own browser, its own file system, and its own session state.
