Definition
Plain language
A design rule that treats leaving out something critical as far worse than keeping something mildly useless — so when in doubt, keep it.
As stated in the literature
An asymmetric-cost heuristic in agent-skill and harness filtering: dropping a genuinely needed component is catastrophic while retaining a marginally harmful one is survivable, so removal is made conservative and defaults to keeping anything that might help, with a fallback to the full set.
Why it matters: It biases pruning toward caution so an agent never accidentally discards a component it can't function without.
For example, when deciding whether to drop a piece of stored advice, the system keeps it if there's any doubt, because losing something truly needed is far worse than carrying a mildly useless tip.
Heard on the show
“It's the parachute principle.”Episode 151 — Why More Experience Made This AI Agent Worse, And How to Fix It