Agents & Tool Use · L4 Foundations

Reliability as a formal object: compounding step error, planning as search, and evaluating agents statistically.

1 step0.95

3 steps0.95^3

5 steps0.95^5

10 steps0.95^10

20 steps0.95^20

End-to-end success rate collapses even with a strong 95%-per-step model.

If each of $n$ steps in an agent's plan succeeds independently with probability $p$ , the probability the entire trajectory succeeds is $p^{n}$ — exponential decay in the number of steps. This single fact explains most observed agent behavior: short, well-verified plans work; long, unchecked ones don't, regardless of how capable the underlying model is.