AI Coding Agents and a Rerun of the Operating-System Wars

Summary

This analysis argues that the software surrounding an AI model may matter as much as the model itself. An agent harness supplies context, exposes tools, controls permissions, and determines how instructions and results move through a workflow. The article compares Claude Code with Windows because third-party extensions and conventions such as plugins, skills, hooks, MCP servers, and CLAUDE.md files can create ecosystem effects that competitors may also need to support. Codex CLI is compared with BSD because its Apache 2.0 code can be modified, reused, and connected to custom providers or local models, even though the project is associated with OpenAI. OpenCode and similar tools are compared with Linux because they place provider choice, local models, and custom endpoints near the center of the user experience. Antigravity is compared with Macintosh because its appeal lies in integrating an editor, terminal, browser, Agent Manager, and Google Workspace services, although some integrations require Developer Preview access, configuration, OAuth, and appropriate permissions. The author argues that “open” has several separate dimensions: inspecting and modifying code, independently distributing extensions, replacing providers, using shared protocols, participating in governance, and carrying workflows to another harness. MCP can improve portability by standardizing connections to tools and data, but it does not make products open source or ensure identical agent behavior. The central risk is that accumulated instructions, integrations, permissions, and review practices may become harder to migrate than the underlying model. The article concludes that teams should evaluate both what attracts them to a platform and how much of their workflow they could take elsewhere.