Why is reliability the core problem in multi-step AI workflows?

A workflow wires models and tools together along defined paths. It is the structured, predictable cousin of an agent: instead of letting the model decide every step, a workflow lays the steps out in advance. Anthropic frames the choice as exactly this, between workflows that follow predefined code paths and agents that direct themselves, and advises starting with the simplest approach that works.^[1]

The reason orchestration exists comes down to multiplication. If a single step succeeds 95% of the time, a chain of twenty such steps does not succeed 95% of the time. It succeeds about 0.95²⁰, which is roughly 36%. Push the per-step reliability to 99% and twenty steps still only reach about 82%.

The curve is unforgiving, and it is why a capable model can still produce an unreliable system once you string enough calls together. A better model shifts the curve — 99% beats 95% — but it never flattens it; even at 99% per step, a hundred-step chain is under even odds. The only other lever is the exponent itself: cut the number of steps, or catch failures before they compound instead of hoping a smarter model avoids them. Every structural pattern below is, at heart, a way to pull one of those two levers.

This is also the structural case for choosing a workflow over a free-running agent. An agent that decides its own next step every turn inherits the full compounding: a bad turn three feeds straight into turn four with nothing between them to catch it. A workflow can insert a check at any fixed point in the chain — validate step three’s output before step four is allowed to run on it — which doesn’t raise any individual step’s reliability, but it resets part of the compounding by catching a bad result before it propagates, instead of letting it silently multiply through however many steps remain.