Request to restore user-controlled Custom Subagent / Agent Orchestration capabilities

What variant of Codex are you using?

App, IDE Extension, CLI,

What feature would you like to see?

1. Subagents are not only about role specialization; their key value is allowing developers to actively manage complex task context consumption through isolated contexts.

1. The new Codex direction reduces user-defined subagents, taking away advanced developers’ control over agent architecture, task delegation, and context management.

1. If ChatGPT/Codex does not want to be an orchestratable primary Agent, does not want to support Agent orchestration workflows, and does not want to be the entry point for development tasks, users will naturally choose other primary Agents to orchestrate ChatGPT/Codex; in that case, ChatGPT/Codex may become only an executor and a participant in someone else’s plan rather than the central platform for software development. Interestingly, OpenAI itself released openai/codex-plugin-cc, which allows Claude Code to invoke Codex, demonstrating that Agent orchestration and multi-Agent collaboration are important requirements in real-world development workflows.

1. Advanced developers are the users most willing to invest significant amounts of money into improving AI workflows, while casual users typically have lower demand for complex Agents and lower willingness to pay. Lowering the barrier may increase user numbers, but it may not necessarily create equivalent business value. OpenAI may currently be prioritizing user growth to strengthen its IPO position, but I do not want advanced user experience to be sacrificed for that goal. I hope OpenAI will reconsider the long-term impact of this strategy.

1. Codex should support both automatic delegation and an Expert Mode, providing a simple experience for general users while preserving custom Agents, models, tools, and workflows for advanced users. Expert Mode should support directly assigning work through @subagent, selecting different models for different subagents, configuring independent tools, permissions, and contexts, and allowing subagents to actively send messages, provide feedback, and dynamically guide the main Agent during task execution. These capabilities existed before, but they now appear to be missing.