That conversation now seems like a special period. All the software program improvement lifecycle, as we all know it, shouldn’t be solely accelerating, however collapsing and being redefined. I not too long ago stated that we should always kill the code review. AI generates code sooner than people can assessment it. PRs pile up or get rubber-stamped. That approval gate now not matches how we do software program engineering now.
But when AI is now producing many of the code and there’s no manner a human may ever learn all of it, how will we share data?
Choices over diffs
My proposal for fixing the code assessment bottleneck was to maneuver the human checkpoint upstream, to reviewing intent, reviewing the contract that code ought to fulfill: specs, plans, constraints and acceptance standards. The identical applies to the knowledge-sharing a part of the assessment course of.
If the workforce opinions intent and acceptance standards earlier than code is generated, the data sharing occurs naturally as a part of planning. You’re not making an attempt to reverse-engineer selections from a 500-line diff. You’re aligning on a handful of key selections earlier than something will get constructed. The reviewer reads 10 strains of selections, not 500 strains of code.