Velocity Is Dead. Measure What Agents Can't Fake

For a year, my team’s velocity meant something. Then it didn’t.

For about a year, my team’s velocity bounced between 5 and 35 story points per release. Predictable enough. Boring, honestly, which is what you want from a number you actually trust.

Then we got a three-month waiver to run a controlled pilot of an AI coding assistant. Eight licenses. Two objectives fixed up front: migrate a base of eighteen microservices to a new platform version, and gain twenty points of unit test coverage. Mandatory human review on every merge request, and every AI-touched MR labeled as such. Nothing loose about it.

The velocity chart exploded. 152 points on a single release. Four to five times normal, without a single extra person on the team. The next month, 49. Then 24. Then back to 5 once the licenses were pulled.

If I had walked that chart into my director’s office as proof the pilot worked, I would have been wrong. Here’s why.

The number that proved the pilot wasn’t on the velocity chart

What actually proved the pilot were the two numbers we picked before we started. The migration hit 100% on all eighteen services. Coverage went from 15% to 37%. Two months instead of the three we were given. Time measured on the targeted tasks: down 92%.

Those are outcomes. Observed, verifiable, and I don’t have to ask the team to confirm them. They’re just true.

The 152 is a side effect. Break the bar down and it’s 45 tasks, 21 bugs, 10 stories. The “normal” month before it was 14 bugs and 8 stories. So the spike is a spike of short tasks the tool knocks out fast. It is not a spike of value reaching a user. Measured in points, that month says “we quintupled.” Measured in outcomes, it says “we did what we promised, faster.” Only the second sentence earns a slide.

And the team learned the real lesson on its own, the one that structures everything I think about this now: the bottleneck moved to review. Writing time collapsed, so we had to harden the validation criteria, because review was the only thing left protecting the outcome. A pilot that doubles output and changes nothing about review is just scheduling next quarter’s accident.

The pilot came back down. Agents never do.

I ran the same experiment on my own projects. Wired agents into code and content production, and inside a month my “things shipped” counter doubled. What those projects actually return didn’t move an inch.

Here’s the part that should worry a leadership team. The pilot inflated velocity for three months and it fell back the moment we removed the tool. Agents inflate it and it never falls back, because the thing doing the inflating became permanent. You don’t get to run the experiment and read the before-and-after. The “after” is now the floor.

Velocity was already a bad measure. I’ve written about that before: it measured agitation, not progress. Agents don’t make it lie, they make it fraudulent without anyone lying. Story points were calibrated on human effort. Delegate the effort and the point measures nothing. Not effort, not value, not risk. The chart climbs, leadership claps, and the only number that genuinely moved is the inference bill. You don’t need Goodhart to explain it. The unit of measure changed its nature under our feet. A team celebrating its velocity in 2026 is celebrating its API invoice.

When the bottleneck moves, the measure has to move with it

Writing used to be the bottleneck. Now it’s review, validation, and getting the thing safely into production. An assisted team produces more diffs than it can honestly read. Measuring production now means measuring the mouth of a funnel nobody’s watching the other end of.

Three numbers survive the arrival of agents. All three sit downstream of the code, because that’s where the bottleneck went.

1. What reached a user, and was observed

Not declared, observed. On my own projects the only measure that has never lied to me is a net that checks the expected deliverables actually exist at their destination. Did the article publish. Did the nightly batch produce its output. When was the last processed record. And it screams when they didn’t. An agent will announce “delivered” ten times a day. The observed output doesn’t care about its mood. Transposed to a team: the sprint report should only contain things you can verify without asking the team.

2. What came back

Escaped defects, regressions, rollbacks. This is the most honest measure of the agent era, precisely because it lags and you can’t inflate it. Code generated and merged too fast doesn’t break in the sprint where you measure it. It breaks two sprints later. Watching velocity is watching the accelerator. Watching escaped defects is watching the rearview mirror. The two don’t tell the same story, and the mirror is the one that’s right.

3. The human review load per delivery

Not to measure people. To measure the flow. Deliveries waiting on review, age of the oldest one, ratio of lines produced to lines a human actually read. If that ratio collapses, the team isn’t accelerating, it’s borrowing. And the interest on that loan gets paid in escaped defects, later, with a penalty.

What a sprint review actually looks like now

What I look at in a sprint review changed its subject. It used to be “how much did we produce.” Now it’s three questions, in order.

What reached someone this sprint, and how do we know? What came back, and why? How many human hours of review did it cost, and is that debt climbing?

None of those has an answer in a velocity chart. If the only thing a team can put on the screen is a line going up, it doesn’t have a sprint review. It has a celebration.

The spine hasn’t changed. Agents just made it visible faster.

My leadership thesis has been the same for years. A crystal-clear objective. Constraints everyone knows: quality, visibility, budget, cross-team collaboration. Full autonomy on the how. Accountability for the results.

Agents hand out enormous autonomy on the how. That is exactly why accountability has to land on outcomes and not on production. When the machine writes the code, “how much code” tells you nothing about who did well.

A team that ships less, where everything that goes out holds, beats a team that produces double and takes back a third of it. That was already true before any of this. Agents just make the gap show up a quarter sooner.