I would like access to a simple capability:
Let normal ChatGPT maintain task state and control external agents.
I don’t need a more powerful model. GPT-5.6 is already enough.
What I need is:
-
ChatGPT can maintain a separate task state outside the normal conversation.
-
ChatGPT can send tasks to different agents and receive their status/results.
-
Agent-to-agent and GPT-to-agent communication can stay in the background.
-
ChatGPT only talks to me when it needs a decision, clarification, or approval.
Why I think this matters:
Users should not have to learn how to use many different agents. ChatGPT already understands the user, so it should choose and manage the right agent for them.
Simple work could be handled by cheaper or local models, while ChatGPT focuses on understanding goals, important decisions, error handling, and final verification.
This could reduce compute and context usage, prevent long conversations from losing the original task, and make agents practical for ordinary users.
In short:
The user talks to ChatGPT.
ChatGPT manages the work.
Agents do the work.
If there is an experimental API, permission, or early-access program that would allow me to test this architecture with normal Chat, I would be very interested in trying it.
A small clarification to my original idea:
I am not asking for the normal Chat to have direct execution permissions.
Chat already gives me a very natural experience: I can request work using normal language, the work can be handled elsewhere, and Chat can give me the status and result. I don’t need to switch interfaces.
I would simply like this connection to become persistent.
The structure could be:
User ↔ Normal Chat ↔ Shared Task State ↔ Work/Manager Chat ↔ Agents
Normal Chat only needs to:
- understand what I want;
- update or read the shared task state;
- send new instructions to the working session;
- receive status, results, and questions.
The Work/Manager Chat keeps the execution permissions and manages the agents.
If the work is progressing normally, I don’t need to see the internal communication.
If the working session needs a decision, Normal Chat asks me. My answer updates the task state, and the working session continues.
This also means my normal conversation does not pollute the working context. I can talk about something completely unrelated while the task keeps its own clean state.
So the capability I am requesting may actually be quite small:
Turn the existing Chat → background work → status/result flow into a persistent, two-way task channel.
No more powerful model is required. The important part is persistent task state and communication between the normal Chat and the working session.
One correction to my diagram:
The user-facing Chat does not need permission to execute anything. I already like the current experience where I can request work in natural language without changing interfaces, and later receive its progress and result.
What I would like to test is simply keeping that relationship alive after the initial handoff. The execution session keeps its own clean context and permissions, while my normal conversation remains separate. When the execution session needs a decision, it asks through my normal Chat; my answer is passed back and the job continues.
In other words, I am not asking to give Chat more execution power. I am asking whether the existing handoff can become a persistent two-way connection between two sessions.
