An agent should not ask a language model
every time it needs to decide.

If you are deciding what to put in front of an agent first, it is one of these three. All three ship today, each is one command, and each number below was measured on records nothing was fitted on.

1. 01

   The guard rail

   shipped

   Every tool call an agent proposes is judged before it runs.

   60.9% cold, 95.7% fittedon records never fitted on

   46 held-out proposals, 26 of them risky. Fitted, 19 of 20 harmless calls run unattended and nothing destructive does. On a separate set of 67 file operations it held all 19 deletions written not to look like a deletion, and all 5 that write a script first and run it after.

   47s

2. 02

   Model routing

   shipped

   Each request goes to the cheapest model that can actually answer it.

   96.2% route, 100% stakeson records never fitted on

   26 held-out requests. An earlier fit scored higher on paper because every training row carried a context clause the eval kept too; with the clause gone it answered the stakes question wrong on seven. This one is measured without it.

   47s

3. 03

   The orchestrator's choice

   shipped

   The decision replaces the language model that picks which agent or tool runs.

   80% next agent, 0 wrong and confidenton records never fitted on

   The question sheet is generated from the MCP tool list, so a new tool is routed the run after it is registered. 25 held-out states. A frontier judge picks the agent more accurately than this does, up to 100%, and never stops to ask: Haiku is confidently wrong on two of the 25 and Sonnet and Opus on one each, where this is confidently wrong on none and hands 14 back.

   50s

Each one is the same call with a different question sheet. Fit it on your own cases, or run it cold and let it abstain until you have any.

The three above are where to start. These are the rest of what ships, in the four places a decision goes in an agent's call path. Pick one and it is a project in a minute, or write your own sheet and ignore the lot.

Starting points, not a catalogue. Every one is the same endpoint with a different question sheet, and a sheet you write yourself gets the same calibration, the same prediction set and the same right to abstain. Where each one sits.