The Cheapest Part of Your Agent Should Decide When the Most Expensive Part Runs

Most “autonomous” agents have a hidden bug in their safety story: the decision about whether to ask a stronger model for review is made by a stronger model. The necessity check costs as much as the thing it gates. So under any cost or latency pressure, the agent learns the same lesson a tired engineer does — stop asking.

We hit this in our own advisor ladder. A risky decision routes to two frontier models (Gemini 3.5 + ChatGPT) over browser lanes we already pay for, escalating to a council on disagreement. Solid — except the moment it flagged risk, it paid the full round-trip. There was no cheap way to ask “does this even need review?” first.

The fix is a zero-budget deterministic floor: a transparent keyword-and-structure classifier that answers “review needed” or “auto-proceed” without touching the network or a single token. Only when it says matters does the agent spend the frontier lane.

The principle: the unit of reliability is the gate, not the model. Make the “should I be careful?” decision the cheapest thing in your system — otherwise you can’t afford to make it, so you quietly stop, and the safety property you designed stops running with it.

Support proof-first autonomy work (sponsorship, not equity): paypal.me/exzilcalanza