A frontier model running an agent loop generates every step itself, and most of those steps are routine. Here a 4B model drafts the run — executing read-only tool calls as it goes, holding the first action that would change anything — and the 397B model reads the whole draft in one pass and returns only how long a prefix it accepts plus a replacement for the first bad step. On two customer-support benchmarks that matches the large model's score at 41% and 54% of its wall time and about a fifth of its calls. The accounting is the honest part: total compute goes up 30–87%, so this buys latency and memory traffic, not FLOPs.