Difference between cursor projects and what codex does?

12d • Chat

My Codex setup has evolved beyond a normal coding assistant. It now works as a background operational coordinator for my daily work, particularly email, quotations and the small decisions that would otherwise constantly interrupt me.

I receive around 120 emails per day. A large percentage is internal chatter or information I need to be aware of but don't need to act on. Others require quotations, replies, follow-ups or specific actions.

Instead of manually processing that stream, I have a coordinator chat in Codex that periodically receives my email and decides what should happen next.

It can classify emails, archive or mark them unread, suggest what I should do, create drafts, send emails when appropriate and delegate work to specialised chats. For example, quotation requests are sent to a dedicated quotation chat with its own strict skill and workflow. When the quotation is finished, that chat communicates the result back to the coordinator, which then tells me that it is ready for review.

The important part is that I remain in the loop without having to operate the system continuously.

Codex's live voice is particularly valuable here. I can be working on something completely different while the coordinator continues processing things in the background. When something genuinely needs my judgement, it can simply talk to me:

«“You received this. I think it should be archived and marked unread. Do you agree?”»

Or:

«“The quotation is ready for you to check.”»

I answer by voice and continue working. Instead of reading and deciding what to do with 120 emails, I mainly deal with the relatively small number of decisions that actually require me.

Why the model matters

This is also why I currently cannot simply replace Codex with Grok Bot, even if Grok can technically perform the same orchestration.

I have tested this.

The problem isn't whether Grok can read an email, delegate to another chat or communicate results back to the coordinator. It can. Most of the architecture works.

The problem is reliability when a specialised workflow contains strict instructions.

My quotation skill, for example, has a defined procedure that needs to be followed consistently. Grok 4.7 usually follows it correctly, but occasionally decides to deviate from the instructions, reinterpret part of the process or take an initiative that wasn't requested.

For casual work, that kind of autonomy can be useful. For quotations, it isn't.

A system processing business operations cannot be judged by whether it behaves correctly most of the time. If a workflow says that steps A → B → C → D must happen in that order, I need the worker to execute A → B → C → D rather than occasionally deciding that A → C → “something clever” is better.

That difference becomes much more important once the system is operating semi-autonomously. I am deliberately not watching every action it takes. The whole point is that I can concentrate on my work and be interrupted only when my judgement is required.

If I have to supervise the agent because I'm unsure whether it followed its skill correctly, the automation loses much of its value.

Continuous learning

The coordinator also has access to a dedicated learning folder.

At the end of each day, I ask it to record what it learned from the day's work. Those observations are written to disk and become persistent knowledge that can improve future behaviour.

So the system isn't just:

Email → AI → Action

It is gradually becoming:

Observe → Decide → Delegate → Act → Ask when necessary → Learn → Improve

The goal isn't to remove me from the process. It's almost the opposite.

I want AI to absorb the enormous volume of low-value attention required by my work while escalating the small number of things where my experience, judgement or approval actually matters.

That's why Codex currently works particularly well for this setup, and why choosing the underlying model isn't simply a question of which model is smartest.

For this workflow, obedience, predictability, tool use, communication between agents and the ability to keep me naturally in the loop are part of the intelligence of the system.

10 comments

Difference between cursor projects and what codex does?