AI-Augmented Verification
Generation capacity is now effectively unbounded. Comprehension capacity is not. Everything in this hub follows from that single asymmetry — and none of it requires changing the framework.
The Constraint Has Moved
For twenty years, the scarce thing on a software team was the ability to produce working software. Almost every practice a Scrum Team uses was shaped, directly or indirectly, by that scarcity. Refinement made work small enough to build. Estimation forecast how much building the team could do. Code review caught the mistakes people make when they build under time pressure.
When agents write production code, that scarcity is gone — and the practices built around it start reporting numbers that mean less than they used to. Throughput rises. Cycle time falls. Every dashboard improves. Meanwhile a second quantity, which nobody was measuring because it had never been scarce, starts running out: the team's capacity to understand what it has accepted.
That is the whole subject of this hub. Not whether to use AI — that argument is over, and it was never very interesting. What matters now is where the bottleneck went, and which Scrum practices need rebuilding around it.
AI doesn't remove the constraint — it moves it from generating work to verifying it. Agents are capacity, not members of the Scrum Team. Capacity does not hold accountability, and no volume of generated work transfers any part of it.
What Actually Broke
Three practices carried an assumption that has quietly stopped being true, and each broke in a different way.
The Definition of Done assumed that whoever wrote the code understood it while writing it. Comprehension was a free by-product of authorship, so no Definition of Done ever needed to ask for it. Now it has to be asked for explicitly, or it does not happen.
Acceptance criteria assumed determinism. Given input X, the system produces output Y, and correctness is a property of a single execution. When the feature is a model, correctness becomes a property of a distribution — and Given-When-Then stops being able to express it.
Code review assumed an author who could answer questions. It was a conversation between two people who had both engaged with the problem. When one side of that conversation is missing, review quietly degrades from “I have verified this” to “I found no reason to object,” which is a different claim entirely.
None of this is a case for changing Scrum. The accountabilities, events, and artifacts hold. What has to change is the substance the team puts inside them.
The Verification Practices
Verification Debt
Work accepted into the Increment that no human on the team genuinely understands. Distinct from technical debt, invisible by default, and discovered by incident rather than by inspection. Three instruments make it visible before it costs you.
Read the Verification Debt Guide →Acceptance Criteria for AI Features
Given-When-Then assumes the same input produces the same output. When it doesn't, criteria that look correct become impossible to execute. What an eval actually is, the four grader types, and why the pass threshold is a Product Owner decision.
Read the Acceptance Criteria Guide →The Golden Set
The artifact where “good” stops being a word and becomes something a team can act on. How a Product Owner builds the first one in an afternoon, how to size it in Refinement, and how to grow it from real production failures.
Read the Golden Set Guide →Reviewing Code Nobody Wrote
Code review assumed an author who understood the change. The comprehension flag, the blast-radius declaration, and how to make “approved on trust” socially safe to admit — because an instrument nobody dares use reports zero forever.
Read the Code Review Guide →A Model Upgrade Is a Breaking Change
Changing a model version, a prompt, or the retrieved context alters behaviour with no code diff to review. Why AI features need regression evals, and how a Scrum Team should plan an upgrade as a Product Backlog item rather than an infrastructure detail.
Read the Model Upgrade Guide →The Retro Question
Standard retrospective prompts collect what people already know they know. Verification debt lives somewhere else. One question, the psychological safety needed to answer it honestly, and what to do with the first answer you get.
Read the Retrospective Guide →The Verification Chain
Each link is a place work can pass through unexamined. The Definition of Done sets the bar, acceptance criteria make that bar testable for probabilistic features, and comprehension at review is what stops the bar from being cleared on paper alone.
AI Scrum Master
Acceleration Course
An Interactive Hands-on course for Scrum Masters, Agile Coaches, and Agile Leaders
VIEW DETAILS