Claude Code の Agent Teams 風機能を自作し Codex でも活用する試み A developer's attempt to recreate Claude Code's Agent Teams-like functionality on their ow…

Anthropic's Claude Code includes an Agent Teams-style mechanism that lets multiple sub-agents collaborate on a task by dividing responsibilities. Inspired by this behavior, the author of this post sets out to build their own equivalent and, importantly, make it usable from OpenAI's Codex CLI as well.

For context, Claude Code lets developers define sub-agents via Markdown files placed in the project, which the main agent can delegate to as needed. This pattern is useful because it isolates context windows and lets each role specialize in a particular kind of task — for example, a reviewer agent, a test-writer agent, or a refactoring agent. Other CLI-based coding tools, including Codex CLI, do not necessarily ship with the same built-in orchestration, so there is a real demand for a portable approach that works across vendors.

The author's approach appears to involve keeping agent definitions and prompts in a shared, tool-agnostic format and then wrapping each CLI with a thin layer that loads those definitions and launches the underlying assistant. In practical terms, this likely means spawning sub-agents as subprocesses or separate sessions, capturing their output, and feeding it back into an orchestrator process that coordinates the overall task. The exact implementation details are still being explored, but the design intent is clearly to avoid duplicating agent specifications per tool.

This effort sits within a broader trend in the agentic coding space. Tools such as Cursor, Cline, Aider, and Codex CLI have all been iterating on their own takes on agent">multi-agent or sub-agent workflows. At the same time, standardization efforts like Anthropic's Subagents specification and the Model Context Protocol (MCP) are pushing toward common interfaces for tool use and agent composition. If those standards mature, the idea of carrying a single agent">multi-agent setup across different coding assistants could become much more practical than it is today.

While this particular project is a personal experiment rather than a production-grade framework, it illustrates a meaningful pattern: developers want to shape their own agent workflows and reuse them regardless of which underlying model or CLI they happen to be using that day. As proprietary agent features proliferate, lightweight portability layers like this may serve as a useful hedge against vendor lock-in, and they hint at what a more interoperable agent ecosystem could look like.