benchmark-hundred-step

measure reliability across 25 to 100 dependent steps

the problem

small per-step mistakes compound until a long agent run ends in the wrong state.

what i built

hundred-step runs chains of dependent actions and measures how reliability falls as the chain grows.