The Number Everyone Wanted
For four years the pitch for AI coding tools has been speed. Write more, faster. This month a Harvard-linked study of more than 700 companies that adopted coding agents put a number on it: roughly 30% more code, about 20% more commits and 23% more pull requests, as Tech Insider reported.
Then came the line that matters. Finished issues and completed epics showed no statistically significant change. More code went into the pipe. The same amount of software came out.
Where the Speed Went
The researchers point at review. Every line an agent writes still has to be read by someone who can say whether it solves the right problem, fits the codebase and doesn’t open a hole. Generation scaled overnight. Review didn’t. The queue simply moved downstream and swallowed the gain.
Builders already act like they know it. In Retool’s 2026 survey of 817 builders, only 8% use AI-generated code without changes, and 44% test it thoroughly before it ships.
Typing Was Never the Hard Part
Software teams were never limited by keystrokes. They were limited by judgment: knowing what to build, proving it works and catching what’s wrong before a customer does. An agent that types faster makes the judgment step more crowded, not less.
More output means more to verify
Self-review is kind to itself
Queues hide in the handoffs
Lines of code are not a business result
Building Around the Bottleneck
Autolace was designed for the second half of the pipe. Review isn’t a queue at the end of the line; it is the line. Tests are compiled from the plan before the code exists. Builders never grade their own work. Reviewers arrive cold, a rival lab’s model reads everything again, and every finding becomes a gate for the next build.
The measure that matters is the one the study found flat: finished, working software. Everything else is volume.
What to Ask Any AI Vendor
How many issues closed, not lines written?
Who reviews the agent’s work, and how?
What happens when a check fails?
What evidence do you hand over at the end?
