AI pipelines generate a growing amount of transient material and intermediate paper trail:
VS Code / Codex threads
ChatGPT chats
document drafts
plan files
stale metadata
intermediate image generations
temporary outputs and side-effect files
AI is understandably conservative about deleting things without explicit authorization, so housekeeping gets pushed back onto its carbon-based counterpart (aka us). At the same time, AI is not always good at inferring the axiological value of an artifact - what is actually worth keeping - without guidance.
It seems like we need a core garbage-collection + triage mechanism.
Periodically, based on user-defined rules, the AI could collect likely-transient artifacts into a new kind of “recycle bin” or staging area. The user could then quickly triage them - keep, archive, promote, or discard.
The important distinction is that the AI would not decide what has value on its own. It would surface likely garbage and make the human decision cheap.
As AI makes it possible to generate artifacts at much greater scale, digital housekeeping probably needs to scale with it.
I believe this is an important matter also when I consider the amount of code being generated in addition to all these artifacts you mentioned. However, it is also a difficult problem and would probably require a lot of time to do the research and adaptation. Thus, I don’t know how much they could prioritize this.
I agree it’s not straightforward, but perhaps we don’t need to solve the hardest version of the problem first.
Trello is an interesting model here. Part of what made it feel so lightweight was that CRUD operations weren’t surrounded by friction and endless confirmations. You dragged something, archived it, restored it, moved on. Changes were cheap and reversible.
OpenAI could experiment with similarly lightweight environments for agents.
For example, an AGENTS.md file could define a “trello-mode”: the codebase is the source of truth, and the agent is explicitly allowed to leave little or no paper trail. Plans, intermediate files, and conversational artifacts are transient by default. Perhaps the main persistent side effect, besides the code itself, is a clearer AGENTS.md for the next agent/thread.
That avoids asking AI to perfectly determine what has value. Instead, the user defines the retention philosophy of the environment up front.
In other words: rather than making AI better at hoarding and then cleaning up, we could give it modes where it simply doesn’t hoard in the first place.