What does cascading error compounding mean for production agent design and reliability budgeting?

**Cascading error** is what happens when a small per-step failure rate adds up across a long chain of steps. An agent that needs every step to succeed is only as reliable as the product of all its step reliabilities. Take the numbers in the question. A 90% per-step rate over 10 steps gives 0.9 multiplied by itself ten times, which is about 0.35. So roughly one run in three finishes correctly, even though 90% sounds high when you hear it alone. The reason is that the steps are linked in series. If any single step fails, the whole task usually fails. The good outcomes have to clear every step in a row, and each step shaves a little off the total. This is why long agent workflows are fragile. Adding more steps does not just add a little risk, it multiplies the risk down. A 20-step version of the same agent would finish only about 12% of the time. The fixes are to make each step more reliable, to use fewer steps, and to add recovery so a failed step can be retried instead of killing the whole run.

**Cascading error** means per-step failures compound multiplicatively, not additively. If each step succeeds independently with probability p, the end to end success over N steps is: $$P_{\text{success}} = p^{N}$$ Plug in the question. At p = 0.9 and N = 10, success is `0.9^10 ≈ 0.35`. So a system that sounds reliable in isolation finishes barely one run in three. The intuition that 90% is good fails badly once steps chain in series. The engineering implication is that per-step targets must be derived from the end to end goal. Invert the formula. To hit 90% end to end over 10 steps you need per-step accuracy near `0.9^(1/10) ≈ 0.989`. That jump from 90% to 98.9% is enormous in practice. Three levers follow. Raise per-step reliability through better prompts, tools, and verification. Shorten the workflow, since a 4-step task at 90% already reaches 66% versus 35% at 10 steps. Add error recovery so a failed step retries instead of aborting, which lifts the effective rate the exponent acts on.