Note
October 18, 2026
2 min read
A Small Recursive Problem
By Cristiano Pierry
AI needs a safe way to work with its own services, with specific permission, visible results, and access I can revoke.
I asked dots, from my phone, to create a new project and start some work on my Mac mini. Then I opened Codex on the Mac mini to find the project.
It wasn’t there.
The snag is that dots can’t add the project, and neither dots nor Codex is allowed to control the ChatGPT app itself. Computer control isn’t disabled in my setup. OpenAI doesn't allow Codex to control itself - for obvious reasons.
I can understand the caution around an AI operating its own host app, especially where permissions are involved. I wouldn’t want “please add this project” to become permission to grant itself whatever access it needs next.
But this also exposes a limitation in something I increasingly want from these tools: AI as the integration layer across my services, apps, and computers. Eventually, that needs to include the services the AI itself runs on.
There ought to be a way for me to authorize a specific action, such as adding a project, and then see the result in the app where I expect to review the work. That permission should have a clear boundary, and I should be able to take it back.
For now, I can ask for work on my Mac mini from my phone. Making that work show up where I can find it still needs some work of its own.
This writing reflects my personal perspectives on product management, AI, and content discovery. It does not represent the official position of my employer or any affiliated organization.