What's this about?
A survey on systems, evaluation and principles for agentic artifact creation. The framing sentence is the useful one: generative models lower the cost of producing drafts and components, and their practical impact increasingly depends on whether those pieces can become complete, dependable deliverables. 52 upvotes.
What is the gap between a draft and a deliverable?
A draft is something plausible in the right shape. A deliverable is something that works, has been checked, fits its context, and that someone will stand behind. The distance between them is where nearly all remaining effort now sits — and it is not proportional to the draft's quality, because a 90%-right artifact can take longer to finish than starting over. This is the same complaint underneath "AI-written code is still your code" on HN today, expressed as a research agenda.
Why does this matter to you?
- A survey arriving means the field considers this a defined problem rather than a complaint, which is progress.
- "Complete and dependable" is the right bar and almost no benchmark measures it.
- It applies to security reports as directly as to code — a finding is a draft until someone has verified it is real.