OpenAI Codex Launches Subagents Mirroring Claude Code Multi-Agent Pattern
The widely supported pattern lets coding agents spawn specialised sub-tasks with custom models and instructions
Codex subagents come in three default flavours: explorer for investigating codebases, worker for running parallel small tasks, and default for general operations. Developers can also define custom agents as TOML files with their own instructions and model assignments — including the lightweight gpt-5.3-codex-spark for speed-critical subtasks.
The pattern is now widely adopted across the coding agent ecosystem. Simon Willison documented implementations across OpenAI Codex, Claude Code, Gemini CLI, and several other platforms, noting the convergence on similar architectures.
Custom agents can be referenced by name in natural language prompts, enabling workflows like: investigate a bug with one agent, trace the code path with another, and implement the fix with a third — all coordinated from a single prompt.
Analysis
Why This Matters
Subagents represent the maturation of coding agents from single-threaded assistants to orchestrated multi-agent systems. This enables tackling complex problems that require different specialisations working in parallel.
Background
Claude Code pioneered the subagent pattern in mainstream coding tools. Codex's adoption confirms it as an industry standard rather than a single-vendor feature.
Key Perspectives
The convergence across multiple platforms suggests the multi-agent pattern is genuinely useful rather than a marketing differentiator. The ability to assign different models to different subtasks is particularly practical.
What to Watch
Whether subagent orchestration becomes the primary way developers interact with coding agents, replacing the single-conversation model that has dominated so far.