When people say AI makes software development faster, I want to know which part got faster. Writing the first version? Fixing a bug? Getting a useful change in front of a customer?
A team can produce more code and still take the same amount of time to release a feature. I’d want to see what changed between the first discussion and the customer using it.
Look at one actual change
I’d take a recent feature and write down what happened from the first discussion to release. How long did it take to decide what to build? How long did implementation take? How long did the change wait for review, integration and testing?
Suppose implementation takes a day and review takes a week. Making implementation twice as fast saves half a day. If the team uses that time to start more features, the reviewer may end up with an even longer queue.
I’d fix the review delay before trying to generate more code. That might mean smaller changes, clearer requirements or giving another person the time and authority to review.
Make the change easier to understand
A reviewer needs to know what the change is supposed to do and how to check it. A large pull request with a vague description makes them reconstruct the intent before they can assess the code.
I’d rather see a small change with a clear reason, the relevant test results and any decisions that still need a person. This helps whether a human or an agent wrote the code.
Some delays need engineering work: a reliable test environment, better regression tests or an easier deployment process. Others need someone to make a decision. The coding tool won’t decide who can approve a release or which customer request should come first.
Measure when the customer gets the result
I’d track elapsed time from an agreed task to a released, checked change. I’d also track rework and failures after release. Compare similar changes before and after introducing the tool or process.
Lines of code and pull-request counts can tell you how much activity there was. To find out whether development improved, I’d check whether customers got useful changes sooner and how much work it took to fix mistakes.
If review accounts for most of the delay, I’d start by changing the review process. If the team repeatedly fixes the same failures after release, I’d examine the requirements and tests. The next change should address the delay or failure the team actually has.
How I check whether an agent actually finished the task · Benchmarks, evals and agent coordination