Feature Request: Allow ChatGPT to Orchestrate External Agents While Preserving Memory and Conversation Context

[rockyinun]

I want to test an architecture where ChatGPT is the long-term cognitive and orchestration layer, while external agents only execute tasks.

The structure is:

User
→ ChatGPT
→ External agents
→ Tools / local models / operating system

The reason I specifically want to use ChatGPT, rather than build a separate API app, is ChatGPT’s long-term memory, conversation context, and understanding of the user’s goals over time.

ChatGPT would:

• understand the user’s intent;

• choose the right agent;

• define the task, constraints, and success criteria;

• intervene only when an agent is stuck or needs permission;

• validate the final result.

The external agents could be a local general-purpose agent, a coding agent, browser/computer-use agent, data-analysis agent, etc.

The communication interface can be very small:

• start_task

• read_task

• steer_task

• stop_task

Normally ChatGPT would not need continuous updates. The execution layer can monitor the agent locally and only notify ChatGPT when:

1. the task is completed;

2. there has been no meaningful progress for a few minutes;

3. a sensitive action requires permission.

My current obstacle is that normal ChatGPT Plus conversations do not seem to support user-defined write-capable MCP/tools.

Using the API allows tool calling, but loses the ChatGPT product-level memory and long-term conversational context that I specifically want to test.

What I need is simply:

Allow a normal ChatGPT conversation to use an external agent as a tool while preserving ChatGPT memory and conversation context.

This could be through full custom MCP, a local tool interface in ChatGPT Desktop, developer mode, or any equivalent mechanism.

I would be interested in testing any beta or developer preview that supports this.

[rockyinun]

@OpenAI_Support — this is closely related to several recent ChatGPT ↔ Codex orchestration requests.

My specific request is for ChatGPT to remain the persistent reasoning and context layer, while delegating execution to Codex or other agents, reviewing the results, and deciding what happens next.

Could this be passed to the relevant product team?

[curt.kennedy]

Have you tried Codex in the ChatGPT app, it sounds similar to what you are thinking.

[rockyinun]

Thanks. Yes, I’ve been using Codex extensively. What I’m proposing is different: I don’t want to switch from ChatGPT to Codex. I want ChatGPT to remain the controller and delegate execution to Codex or other agents, while keeping the conversation context and reviewing the results. In other words: ChatGPT thinks and coordinates; agents execute.