RedHub AI
Quick Kit · The Math
How the math works
Why a "reliable" agent quietly fails
Every step in the chain looks fine on its own. Multiply them and the true end-to-end number is the one nobody watches.
The inputs (one refund-processing agent)
Chain length
8 steps
start to finish
Per-step reliability
85%
every step
Naive guess
~85%
if you don't compound
Long-chain floor
70%
gate at 6+ steps
0.85 ^8 × 100
= 0.85 ^2 = 0.72 → ^4 = 0.52 → ^8 = 0.272 = 27.2% end-to-end
True end-to-end · 8 steps
27.2%succeed
Runs clean start to finish ~1 time in 4
Each step is 85% reliable. The whole chain is worse than a coin flip.
Eight steps that are each 85% reliable run clean end-to-end just 27% of the time. The kit computes true end-to-end reliability for every agent and flags the long chains before they quietly fail.
Expected end-to-end = per-step reliability raised to the number of steps. A monitoring aid, not a guarantee of agent behavior.