Setup travels, ability does not
You cannot copy a strong engineer into four teams. You can copy the checks they run, the context they supply and the way they frame a task, and all three are properties of setup.
Your best team's way of working, taught to the rest.
Give two teams the same AI tools and you still get different quality back. The difference isn't the model — it's the craft of steering it: how a run is scheduled and shaped, what it's handed before it starts, and how clearly the work is asked for. That craft can be measured, copied and made standard.
Cost per merged PR by team
Data$1.06
Platform$1.21
Payments$2.94
Same tools, same access
Getting good output back is a craft, not a setting.
Agent-steering is that craft. It has three parts — each one varies between teams with identical tooling, and each one can be fixed for everybody at once.
Prompt Engineering
How the work is described. The same model returns usable code or a near-miss depending on how the request was framed, and that framing is a skill your org can raise.
Loop Engineering
Work an agent runs on its own — scheduled jobs, repair cycles, multi-step passes. When nobody is watching a run, cost and quality both drift quietly.
Graph Engineering
How a multi-step run is wired — which steps actually depend on each other and which just happen to run in the order they were written. Untangling the two frees up speed nobody has claimed.
How a difference between two teams becomes a standard for all of them.
The same four moves apply to every part of that craft, which is why they sit on one layer rather than three unrelated pages.
See the spread
Today every team looks broadly the same in your reporting, so a real difference in effectiveness has nowhere to show up.
What Tetriz does here
- One number, comparable across teams
Cost per merged PR, built from actual sessions and the code they produced.
- Held to the same kind of work
Reported by work type, so a team on infrastructure is not measured against one shipping small features.
Cost per merged PR by team
Data$1.06
Platform$1.21
Payments$2.94
Find where the return is
The instinct is to make the strongest team stronger. The arithmetic almost always says the opposite.
What Tetriz does here
- The distance from weakest to median
Larger than the distance from your best team to perfect, and it closes without new tooling.
- Ranked by what moving it is worth
So attention goes where it pays rather than where the noise is.
Gap to the cheapest team
| Team | Cost / PR | Gap to best |
|---|---|---|
| Data | $1.06 | — |
| Platform | $1.21 | 0.15 |
| Payments | $2.94 | 1.88 |
Name what the strong team does
Without this the answer is always seniority or luck, and neither of those can be handed to another team.
What Tetriz does here
- Traced to how the work was steered
How the work was aimed, what it carried, how it was asked for.
- Specific enough to copy
A named check, a named piece of context, a named shape of request — not a principle.
Practice comparison · strongest vs weakest
| Strongest | Weakest | |
|---|---|---|
| Checks on a run | 3.1 | 0.7 |
| Context complete | 91% | 54% |
| Reworked | 9% | 34% |
Make it arrive by default
A practice that depends on being remembered stops happening the week someone is busy.
What Tetriz does here
- Published as setup, not advice
It becomes part of the environment and the knowledge every team already has.
- There on the first session
A new team gets it without a rollout, and a new joiner without a mentor.
Org standard rollout
On standard
9 of 11
Teams
3 of 3
Days to adopt
1
Why raising your weakest team beats sharpening your best.
Three reasons the return sits at the bottom of the distribution rather than the top.
The distance is bigger down there
Your strongest team is already near what good looks like, so there is little left to win. The gap between your weakest team and your median is usually several times larger.
And it holds better than habit
Something published into the environment keeps applying long after the quarter that introduced it and the person who championed it moves on. When it does drift, that drift is what the harness layer is built to surface — not something you find out about by accident.