Definition
Plain language
A grid showing, for each piece of an AI agent's stored advice, which kinds of tasks it actually helps on and which it quietly hurts.
As stated in the literature
A skill-by-task grid whose cells estimate each skill's causal effect on each task type via randomized inclusion masks, exposing that most skills are causally heterogeneous — helping on some tasks and harming on others while netting near zero on average.
Why it matters: It exposes that stored advice is rarely universally good, so blindly piling up skills can quietly degrade performance unless you know which task each one actually helps.
For example, the grid might reveal that a 'always double-check arithmetic' tip boosts the agent on math problems but slows it down and hurts it on quick lookup tasks.
Heard on the show
“And what falls out is what they call an attribution matrix — that green-and-red grid.”Episode 151 — Why More Experience Made This AI Agent Worse, And How to Fix It