OpenAI Codex subagents reach GA — manager-worker model, up to 8 parallel
Codex's subagent feature went GA on March 14, 2026 with a manager-worker model supporting up to 8 parallel workers per task. As of May 2026 Codex still holds the top spot on the most-cited coding benchmark.
The shape: a manager agent decomposes a task into subtasks, dispatches them to worker agents, and integrates results. Up to 8 workers can run simultaneously. Each worker has its own context and tool budget; the manager owns the integration step and the user-facing reasoning trace.
This is the same architectural pattern Anthropic ships in Claude Code (the Task tool with specialized subagent types) and Cursor 3 (Agents Window). Three independent convergences on manager-worker-up-to-8 suggest this is the local optimum for the current generation of agent runtimes — enough parallelism to amortize per-task fixed costs without the coordination overhead growing combinatorially.