Reliability Levers

Every dependent transformation multiplies risk. An agent run combines model inference, retrieved context, tool results, external systems, intermediate artifacts, and final verification. Each output can become input to the next step, so a small loss early in the chain can shape everything downstream.

A useful first approximation is:

R(n) = p^n

If every step succeeds with probability p, then all n dependent steps succeed with probability p^n. At p = 0.95, a 20-step chain succeeds only about 36% of the time.

This is a smooth simplifying model, not a forecast. It assumes equally difficult steps, stable per-step reliability, independent failures, and a strict definition of full-task success. Real agent work violates all four assumptions. The model is still useful because it makes one operational constraint obvious: adding dependent transformations cannot improve end-to-end reliability on its own.

Reliability levers

Dependent transformations multiply risk

Under the simplified assumptions, full-task reliability follows R(n) = 0.95ⁿ. Representative values show how a strong per-step rate compounds across a chain.

Real Decay Is Uneven

Production runs do not lose reliability on a clean exponential schedule. Routine transformations may preserve state for several steps. An ambiguous decision may cause a small loss. A wrong premise can then produce a cascading drop because later steps inherit it.

Those failures are correlated. Once generated output becomes working context, the next step does not start fresh: it reasons from both the useful signal and any distortion already present. A wrong import can lead to the wrong API, then to a coherent fix for the wrong problem.

The smooth baseline and an uneven run describe different things. The baseline exposes the cost of dependency length. The uneven trajectory shows why where a loss occurs—and whether it propagates—matters as much as the number of steps.

Reliability levers

Real runs plateau, drift, and cascade

A smooth baseline is useful for reasoning about chain length. Real runs are uneven: routine steps may plateau, ambiguity causes local loss, and a sticky premise can correlate failures across later steps.

This propagation tendency is failure stickiness: the chance that a failed step makes a later dependent step more likely to fail. Reliable workflows control both baseline step quality and the distance bad state can travel.

Next: Reliability: Context Quality