AI agents now write, review, and refactor real code. That is not a prediction, it is how a growing number of teams already work, mine included. It is also quietly breaking the way most teams measure engineering velocity, because nearly every metric we reached for in the past was a proxy for how much code humans produced. When producing code gets dramatically cheaper, those proxies stop measuring anything useful, and a few of them start actively lying to you.
The old metrics were always shaky. Now they are dangerous.
Lines of code was never a good measure of progress, and in a world where an agent can generate a thousand lines before lunch, it is noise. Story points and velocity charts inflate when output is cheap, so a rising line can mean the team is shipping value or simply that the agents are busy. Pull request counts are easy to game and easier still to flood. Every one of these metrics shares the same flaw. It counts output, and output is now the cheap part.
If you reward output in the age of agentic programming, you will get a great deal of output. Much of it will be slop, and someone will have to clean it up.
Measure outcomes, not output
The fix is to measure the things that still cost something real: validated value delivered, and the health of the system that delivers it.
The delivery metrics that have held up for years are a good starting point, because they were never about volume. How quickly does an idea reach production. How often do you deploy. How often does a change cause a failure. How fast do you recover when it does. These resist gaming because they track flow and stability, not effort, and they map cleanly onto outcomes a business actually cares about.
To those I would add the measures that matter specifically when agents are in the loop. How much delivered work survives review without rework. How often defects escape to customers. How much of your shipped change turns into value rather than churn. Those tell you whether the speed is real.
The bottleneck moved, so the measurement should too
Here is the deeper shift. When a machine can produce code quickly, the constraint on a team is no longer typing. It moves to the human work agents cannot do well on their own: deciding what to build, specifying it clearly, and reviewing and verifying what comes back. The way I run my own teams, the work is directed, reviewed, and gated, with senior judgment at the front and the end of every change. That is where quality is now won or lost. I use a simple thought experiment to make this concrete with teams.
So measure that work. How long is the cycle from a well-formed idea to validated production. Where does review become the bottleneck. How much rework are you absorbing. These questions point at the real constraint, and improving them actually makes the team faster rather than just busier.
What this means for leaders
The teams that win with agentic programming will not be the ones that generate the most code. They will be the ones with the best direction and the most disciplined review, supported by infrastructure that lets them move quickly without lowering the bar. Cheap code raises the value of good judgment, it does not lower it. Your metrics should reflect that, or they will quietly steer your team toward producing more and accomplishing less.
Helping teams adopt agentic workflows, and measure them in a way that rewards outcomes instead of output, is a growing part of what I do as a fractional CTO with Artificer Innovations. If your team is wrestling with this, let’s talk.