While working on a more complex task with DALL·E and ChatGPT (creating a photorealistic scene of a mole as a Prague metro driver), I discovered that GPT is already almost capable of acting as a fully autonomous visual agent.
In several cases, GPT was able to:
- Generate an image,
- Analyze the result autonomously,
- Adjust the prompt without user input,
- Repeat the process intelligently.
Surprisingly, in some sessions, the usual user confirmation step (after every single generation) was bypassed, possibly due to how sequential tasks were interpreted. This accidental behavior effectively demonstrated that GPT can already function as a simple iterative agent.
Suggestion:
It would be great if users could optionally allow GPT to:
- Perform a specified number of autonomous iterations (e.g., up to 20),
- Use low-resolution images for intermediate steps to save computing time,
- Automatically switch to full-resolution output after a satisfactory prompt is found.
This would save enormous time, reduce friction for creators, and make better use of computing resources.
My concern is that such autonomous behavior might be further restricted by future safety updates, but I personally see it as a huge opportunity for OpenAI to lead in practical AI agent development.
Is OpenAI considering unlocking such agent-like workflows for power users?
Thanks for considering this idea!