AI output volume is up and the dashboard says adoption works. Escaped defects and rework say otherwise. The gap is a standard your team never rewrote: what "done" certifies once an agent wrote the code.
Output is up but the metric has not moved. The reflex is to add a standup. Often the stall is an under-defined result, not a coordination gap. The shape of the stall tells you which lever to pull, and they are not interchangeable.
Self-report drifts upward and license counts measure procurement, not practice. The signal that holds up is the artifact each role produces every week, read on a ladder from exists to outcome.
Skills, subagents, hooks, MCP. Most leaders treat them as one pile of jargon and let whoever wired the agent that morning decide. Each answers a different question, and the question is yours.
The model writes the function in four seconds and the team still does not move. AI removed the cost of writing code and left the cost of specifying, governing, reviewing, and remembering it untouched. Six artifact classes are what close that gap.
The cut was supposed to prove AI was working. The quiet rehiring is the proof that something else was. AI takes the task; it does not inherit the judgment, ownership, and review standard the seat also held.
The survey says the team is well along. The license dashboard says every seat is active. Then you open last sprint's pull request and it reads exactly like a year ago. Here is how to profile what actually changed, role by role, from the artifacts.
A background agent opened a PR at 2 a.m. and the gate let it through, because the gate was built for a worker who could be reasoned with. The DevOps job is no longer writing infrastructure. It is owning the bounded environment agents act inside.
Connecting an MCP server is not enabling a feature. It is taking on a code dependency that holds execution and data rights nobody reviewed. You already govern packages and APIs as a supply chain. Your MCP fleet is the same risk under a new name.
For two years teams filed prompt injection under "someone else's problem." That filing expired. The moment an agent reads content it did not author, every author of that content can rewrite its instructions.
An AI agent is not a smarter chatbot. It is a new reader of your system of record, and it can only read what you have made authoritative, current, and governed. The Klarna case is the warning: govern the substrate before you scale the reader.
Slow human execution used to absorb your org's ambiguity before anyone noticed. AI compresses that buffer, so thin specs, unclear ownership, and broken handoffs surface at speed. The fix is the operating model, not another tool.