What did Claude Code code?

The middle layer of software engineering is changing. It is becoming easier to ship than ever. Claude Code makes 10x engineers. Sub-agents. Max plans. Auto-accept all changes. But what shipped? Did it fulfil the requirements? There are tons of tools available to automatically review code for syntax, bugs and test coverage. But did the release deliver the feature the customer requested? Did it fix the bug? Maybe it did, how do you loop it back?

The before

Software work has traditionally started with an idea making its way through formal docs like PRDs, until eventually someone opens their IDE, writes the code, and ships it.

Idea -> Work -> Delivery

This is changing

Deciding what actually needs to be built remains the most important step. But now, AI coding workflows can deliver working code from goals, context and tasks. Making the middle layer thinner, more automated, with less time spent translating intent into implementation. Great, you can now ship so much faster!

But...

With AI coding making the middle step faster, more opaque, and less supervised, it becomes increasingly important to verify that what you wanted to build, was what actually ended up being built. Many companies were struggling with this even before AI coding. Traditionally this work has fallen to PMs, verifying releases against stories, updating roadmaps, and writing changelogs. But at the rate we're are able to produce code in this paradigm - a rate likely to increase - this becomes infeasible. When you can kick off a bunch of Claude Code sub-agents from inside of a Linear ticket, instead of carefully planning and spacing out work in 2-week sprints, where do you think the bottleneck will reveal itself? The tooling and workflows at either end of the cycle need to improve and change to handle this, blend in with the overall process, and most importantly, loop back.

AI-native companies, how are you handling this?