Your point that a strong agent can build the wrong thing more efficiently is exactly why I see specifications as context infrastructure, not preamble. One extension I would add is provenance: every acceptance criterion should retain the decision, source, or user signal that created it. That gives the PM agent something to challenge, the engineer a boundary to respect, and QA a reason behind the check. Otherwise a well-groomed issue can still preserve a stale assumption with impressive precision. The graph becomes much safer when its handoffs carry both requirements and their evidence.
This was really cool, thank you! As someone who worked as a software engineer before, it's incredible how fast you can get a project set up with this AI-native workflow. I can also see it removing many of the long grooming meetings with the product team.
Your point that a strong agent can build the wrong thing more efficiently is exactly why I see specifications as context infrastructure, not preamble. One extension I would add is provenance: every acceptance criterion should retain the decision, source, or user signal that created it. That gives the PM agent something to challenge, the engineer a boundary to respect, and QA a reason behind the check. Otherwise a well-groomed issue can still preserve a stale assumption with impressive precision. The graph becomes much safer when its handoffs carry both requirements and their evidence.
This was really cool, thank you! As someone who worked as a software engineer before, it's incredible how fast you can get a project set up with this AI-native workflow. I can also see it removing many of the long grooming meetings with the product team.
Very interesting, and thanks for this!